Multiscale Analysis of NMC811 Lithium-Ion Battery Cathodes: Electron Microscopy and Machine Learning Approaches PhD Thesis May Ching Lai University of Cambridge This thesis is submitted for the degree of: Doctor of Philosophy Gonville and Caius College May 2025 Declaration This thesis is the result of my own work and includes nothing which is the outcome of work done in collaboration except as declared in the preface and specified in the text. It is not substantially the same as any work that has already been submitted, or is being concurrently submitted, for any degree, diploma or other qualification at the University of Cambridge or any other University or similar institution except as declared in the preface and specified in the text. It does not exceed the prescribed word limit for the relevant Degree Committee. May Ching Lai May 2026 Abstract The rapid growth of electric vehicles has increased demand for high-energy, long-life lithium- ion batteries (LIBs). Ni-rich layered oxides such as LiNi0.8Mn0.1Co0.1O2 (NMC811) are leading cathode candidates, but their performance is limited by degradation that initiates at the nanoscale, including surface reconstruction, microcracking and cation mixing. Understanding how these processes develop, and how they are influenced by electrode processing and operating history, is essential for improving durability. This thesis develops and applies quantitative electron microscopy and data-driven analysis methods to connect processing, 3- dimensional (3D) microstructure and local redox evolution across multiple length scales. Phase-resolved 3D FIB-SEM tomography combined with machine-learning segmentation is first used to quantify microstructure and transport in calendered single-crystal (SC) and polycrystalline (PC) NMC811 electrodes. A multi-method tortuosity framework: graph-based analysis, TauFactor simulations and symmetric-cell electrochemical impedance spectroscopy (EIS), shows that SC electrodes preserve structural integrity under heavy calendering and can maintain through-plane transport. In contrast, PC electrodes are more susceptible to intergranular damage and loss of connected pathways. These trends support distinct processing targets: heavier calendering for SC electrodes (~ 25 % porosity) and moderate calendering for PC electrodes (~35 % porosity). The thesis then develops a multi-Gaussian fitting workflow for STEM-EELS that enables robust extraction of transition-metal L-edge metrics and spatial mapping of oxidation-state gradients on non-monochromated systems. Applied to cycled NMC811, the approach links voltage-time history to protocol-dependent surface redox evolution, showing that prolonged high-voltage exposure is associated with thicker degraded surface layers (~10 to 20 nm) and stronger oxygen-loss signatures. Finally, a convolutional neural network achieves 98.8 % accuracy in classifying Ni oxidation states from EELS spectra, enabling rapid nanoscale redox mapping with quantified confidence. Together, these methods provide a reproducible framework for diagnosing degradation and guiding microstructural and protocol optimisation in Ni-rich cathodes. Acknowledgements This PhD would not have been possible without the support, encouragement, and inspiration of many people who have accompanied me throughout this journey. First and foremost, I would like to express my gratitude to Professor Cate Ducati, my supervisor, for her boundless enthusiasm, belief in my potential, and for always making time for thought-provoking conversations even at her busiest. Her dedication to extracting every last piece of information an atom could offer, and the freedom she gave me to explore my interests, from training on six electron microscopes to reshaping the MPhil FIB-SEM and SEM demonstrations around my interest for batteries, shaped the way I think as a scientist. I am grateful to the Faraday Institution - Degradation Project for funding my research and providing opportunities to connect with a wider scientific network. I also appreciate the support from NanoDTC and Gonville & Caius College, which enabled me to attend conferences that greatly contributed to my scientific and personal development. I’m especially grateful to Dr Giorgio Divitini, who acted as a de-facto second supervisor when I joined the group during the height of COVID-19 in October 2020. From guiding me through high resolution TEM imaging and early coding struggles to involving me in numerous collaborations - ironically, none battery-related, his support has been invaluable throughout my PhD journey. To Dr Jędrzej Morzy, thank you for your early mentorship and for sharing your deep knowledge of batteries. Even after relocating to Switzerland, your continued input and regular meetings helped shape the direction of my research and kept me on track. To the people who trained me on electron microscope and allowing me to be a self-reliant microscopist. To Dr Simon Fairclough, my office neighbour, thank you for patiently training me on spectral acquisition (even on a Sunday!), for bouncing around ideas to improve data quality, and for the laughs and banter that made even failed experiments feel lighter. To Dr Petr Vacek, thank you for always being willing to troubleshoot TEM issues and for your dedication to achieving higher-quality results together. To Chris Dolan and Simon Griggs, thank you for bravely training me after a nine-month pandemic hiatus on TEM and FIB-SEM. I’ve been fortunate to work with excellent collaborators throughout this PhD. Dr Kumar Raju, thank you for consistently providing samples and stepping in with electrochemical measurements that were crucial for validating my results. I’m also grateful to Dr Alexander Dimitrijevic for sharing cycled samples under different protocols and explaining the electrochemical context, which strengthened my data interpretation. A special thanks to Dr Daniel del Pozo Bueno for introducing me to machine learning in EELS and for the many hours spent during a conference, which clarified key challenges and sparked my ongoing interest in the field. I’ve also been incredibly lucky to have wonderful groupmates. Dr George Lewis and Dr Joonatan Laulainen, thank you for the warm welcome when I first joined the group, and for the socially distanced walks and picnics that helped during lockdown. Dr Jordi Ferrer Orri, thank you for being the EM group’s social heartbeat - sharing Jupyter tips, squash games, and Stranks group collaborations. Your mix of science and humour made microscope hours enjoyable. Helen Leung, thank you for your support, from lab frustrations to chaotic conferences adventures. And to Professor Charles Footer, who reappeared in my final year, thank you for always reminding me that nothing is impossible, and for your thoughtful feedback on my competition materials. To the rest of the group, thank you for creating a collaborative and welcoming environment, it’s been a privilege to be part of this team. To my friends in Cambridge: Chen Gang, thank you for being a supportive and intellectually stimulating housemate, always challenging ideas and keeping things lively. Ellie, thank you for expanding my horizons, from golf to travels. Caleb, Shafiq, Scott, and Sara, thank you for welcoming me into your business school community and helping me reconnect with normality post-lockdown in my second year. Calista and Sean, thank you for the memories we made in my third year and the continued friendship through video calls and messages. To my long-distance friends, thank you for cheering me on from afar. Adeline and Chai Yen, I am grateful for a friendship that stretches all the way back to childhood. Stephanie, Julia, Julie, and Rebecca Batcup, your continued support since high school, whether through quick meetups during layovers or messages, has meant so much. Sher Lynn, Katrina, Ruth, Siu Ping and Carmen - thank you for the social support, the listening ears, and the pep talks that guided me through rough patches. Deeva and Lexi, thank you for the friendship we’ve shared since our Master’s days - for the support, travels, and all the in-between moments. I feel especially lucky to have had Kelvin Chan, for being by my side during the final stretch of this PhD. Your support, both scientific and emotional, made a meaningful difference. From patiently listening to me rehearse my STEM for Britain speech to helping with coding challenges, you have been a reliable sounding board and a steady source of support. Finally and most importantly, to my family. A huge thank you to my parents for their financial support and for instilling in us the drive to pursue education to its highest level. To Dr May Ling, thank you for being a trailblazer, the first in the family to come to Cambridge for PhD. To Dr May Hsim, sharing half our PhD in the same department, from lunchtime chats and spontaneous dinners to travel plans, kept us both sane and made the experience all the more meaningful. And to my brother, Kien Hsin, thanks for holding the fort at home with your golden hand, and for cheering me on from afar. To everyone who has shared a moment, an idea, a smile, or a word of encouragement, thank you. This thesis is a reflection not only of my work but also of the support and inspiration I’ve received from all of you. List of Publications 1. Raju K, Price SWT, Merryweather AJ, Radić A, Lai MC et al. Enhancing power density and cycle life of NMC811 battery cathodes via combined dense calendering and laser patterning, Energy & Environmental Science 19, 1341–1351 (2026). https://doi.org/10.1039/D5EE06773A 2. Raju K, Wheatcroft L, Lai MC, et al. Influence of Cathode Calendering Density on the Cycling Stability of Li-Ion Batteries Using NMC811 Single or Poly Crystalline Particles. Journal of The Electrochemical Society 171, 080519 (2024). https://doi.org/10.1149/1945- 7111/ad6378 3. Morzy JK, Dose WM, Lai MC et al. Origins and importance of intragranular cracking in layered lithium transition metal oxide cathodes. ACS Applied Energy Materials 7, 3945-3956 (2024). https://doi.org/10.1021/acsaem.4c00279 4. Ooi ZY, Nie S, Vega G, Lai MC et al. Resonant Cavity Effect for Spectrally Tunable and Efficient Narrowband Perovskite Photodetectors. ACS Photonics 12, 4119–4129 (2025). https://doi.org/10.1021/acsphotonics.4c01942 5. Yao C, Leahu G, Holicky M, Liu S, Fenech-Salerno B, Lai MC, et al. Thermally Conductive Hexagonal Boron Nitride/Polymer Composites for Efficient Heat Transport. Advance Functional Materials 34, 2405235 (2024). https://doi.org/10.1002/adfm.202405235 6. Sun, Y., Ge L, Dai L, Cho C, Ferrer Orri J, Ji K, Zelewski SJ, Liu Y, Mirabelli AJ, Zhang Y, Huang JY, Wang Y, Gong K, Lai MC et al. Bright and stable perovskite light-emitting diodes in the near-infrared range. Nature 615, 830-835 (2023). https://doi.org/10.17863/CAM.92711 7. Torsello D, Casalegno V, Divitini G, Ghigo G; Gerbaldo R; Fracasso M; D’Isanto F, MC Lai et al. Triple ion beam irradiation of glass-ceramic materials for nuclear fusion technology. Journal of Nuclear Materials 567, 153783 (2022). https://doi.org/10.1016/j.jnucmat.2022.153783 https://doi.org/10.1039/D5EE06773A https://doi.org/10.1149/1945-7111/ad6378 https://doi.org/10.1149/1945-7111/ad6378 https://doi.org/10.1021/acsaem.4c00279 https://doi.org/10.1021/acsphotonics.4c01942 https://doi.org/10.1016/j.jnucmat.2022.153783 Selected Conference Presentations 1. STEM for Britain Competition (Bronze), 2025. Impact of Calendering on Tortuosity and Transport Properties in Single-crystal and Polycrystalline LiNi0.8Mn0.1Co0.1O2 cathodes via 3D imaging 2. EDGE Conference, Canada, 2024. Advancing Nickel-Rich Cathodes Structural Transformations and Stability Insights through STEM-EELS 3. International Microscopy Congress, 2023. The Role of Calendering in Optimising Electrochemical Performance in Ni-based Battery Materials using FIB SEM Tomography. 4. International Microscopy Congress, 2023. Investigating the degradation and structural changes of Li(Ni0.8Mn0.1Co0.1)O2 cathodes in Lithium-ion batteries using advanced TEM techniques 5. Microscience Microscopy Congress, 2023. Investigating Cracking in Ni-rich Transition Metal Oxide Cathodes 6. Institute of Physics’s Electron Microscopy and Analysis Group Conference Competition (Poster Winner), 2022. Electron Microscopy Studies on Concentration Gradient Ni-rich Cathodes for Lithium-ion Batteries. 7. Armourers and Brasiers Cambridge Forum (Poster Winner), 2022. Electron Microscopy Characterisation of Glass-Ceramic Materials for Nuclear Fusion Technology 8. Royal Microscopical Society Beginners’ Competition (Presentation Winner), 2021. Combining Various TEM Techniques to Characterise Glass-Ceramics for Nuclear Fusion Technology List of Abbreviations Batteries EV Electric Vehicle LIB Lithium-ion Battery NMC LiNi1-x-yMnxCoyO2 Layered lithium-transition metal oxide containing Ni, Mn and Co. Often followed by a set of 3 integers to indicate the exact ratio of the transition metals, e.g. NMC811 stands for LiNi0.8Mn0.1Co0.1O2 LCO LiCoO2 (Lithium Cobalt Oxide) LFP LiFePO4 (Lithium Iron Phosphate) LMO LiMn2O4 (Lithium Manganese Oxide) LNO LiNiO2 (Lithium Nickel Oxide) NCA LiNi1-x-y CoxAlxO2 (Lithium Nickel Cobalt Aluminum Oxide) RSL Rock-Salt Layer PC Polycrystalline SC Single Crystal CBD Carbon-Binder Domain UC Uncalendered Electrochemical Terms SOC State of Charge DOD Depth of Discharge CV Constant Voltage CC Constant Current CCCV Constant Current-Constant Voltage EIS Electrochemical Impedance Spectroscopy GIIT Galvanostatic Intermittent Titration Technique HPPC Hybrid Pulse Power Characterization ASI Area-Specific Impedance Nₘ MacMullin Number (ratio of bulk electrolyte conductivity to effective conductivity) 𝝉 Tortuosity 𝜺 Porosity 𝑹𝜴 Ohmic Resistance 𝑹𝑪𝑻 Charge Transfer Resistance 𝑹𝒊𝒐𝒏 Ionic Resistance TLM Transmission Line Model Specific Energy Energy per unit volume (Wh kg-1) Specific Power Power per unit mass (W kg-1) Vcell Cell voltage (V) Ccell Cell capacity (Ah or mAh g-1) C-rate Rate of (dis)charge, expressed as xC which stands for one (dis)charge half-cycle per 1/x hours (eg: C/20 means that each half cycle takes 20 hours) SOC State of charge Electron Microscopy and Data Analysis Term FIB-SEM Focused Ion Beam-Scanning Electron Microscopy TEM Transmission Electron Microscopy STEM Scanning Transmission Electron Microscopy EELS Electron Energy Loss Spectroscopy HAADF High-Angle Annular Dark-Field ELNES Energy Loss Near Edge Structure XAS X-ray Absorption Spectroscopy TEY Total Electron Yield FY Fluorescence Yield STA Shine-Through Artifact Z-contrast Atomic number contrast t/λ Thickness to mean free path ratio ROI Region of Interest CNN Convolutional Neural Network SVM Support Vector Machine FFT Fast Fourier Transform L3/L2 Ratio of L3 to L3 edge intensities (transition metal white-line ratio) ΔE Energy difference between pre-peak and main peak in oxygen K-edge SNR Signal-to-Noise Ratio Table of Contents i Table of Contents Chapter 1 Introduction 1 Chapter 2 Cathode Materials for Lithium-ion Batteries: Evolution, Challenges, and Characterisation 6 2.1 Fundamentals of Lithium-ion Battery Operation ...................................................... 6 2.2 Evolution of Cathode Materials ................................................................................... 9 2.2.1 Early Cathode Materials ..................................................................................... 10 2.2.2 Development of Mixed Transition Metal Oxide .............................................. 11 2.2.3 Emergence of Ni-rich NMC Cathodes .............................................................. 11 2.3 Degradation of NMC811 ............................................................................................. 13 2.3.1 Bulk Phase Transformation at High Voltage ................................................... 13 2.3.2 Microcracks ........................................................................................................... 14 2.3.3 Ni/Li site exchange .............................................................................................. 15 2.3.4 Surface Reconstruction ........................................................................................ 16 2.3.5 Surface-bulk modulation .................................................................................... 18 2.3.6 Modification Strategies and Single-Crystal versus Polycrystalline Morphologies ....................................................................................................................... 20 2.4 Importance of Electron Microscopy in Battery Cathode Studies ........................... 22 2.4.1 High Resolution Imaging ................................................................................... 23 2.4.2 Spectroscopic Information: EELS ...................................................................... 23 2.4.3 Data Analysis: Machine Learning...................................................................... 26 2.4.4 3-Dimensional Tomography: FIB-SEM ............................................................. 29 2.5 Summary ........................................................................................................................ 31 Chapter 3 Experimental Methods 33 3.1 Electron Microscopy..................................................................................................... 34 ii Table of Contents 3.1.1 Electron Beam Interaction: Elastic vs Inelastic Scattering .............................. 34 3.2 Transmission Electron Microscopy ............................................................................ 36 3.2.1 Electron Gun ......................................................................................................... 36 3.2.2 Condenser Lens .................................................................................................... 38 3.2.3 Objective Lens and Imaging System ................................................................. 39 3.2.4 Detectors................................................................................................................ 40 3.3 Scanning Transmission Electron Microscopy ........................................................... 41 3.3.1 Historical Development and Principles ............................................................ 41 3.3.2 Probe Formation and Beam Scanning ............................................................... 41 3.3.3 Detection system and signal collections ........................................................... 42 3.3.4 Sampling in STEM ............................................................................................... 44 3.4 Electron Energy Loss Spectroscopy ........................................................................... 45 3.4.1 Multiple Scattering Considerations ................................................................... 46 3.4.2 EELS Acquisition in this Thesis ......................................................................... 50 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) ............................ 54 3.5.1 Instrumentation and Operating Principles ...................................................... 54 3.5.2 TEM Lamella Preparation ................................................................................... 56 3.5.3 Modified Mounting Approach for Ultrathin Lamellae for EELS .................. 57 3.5.4 FIB-SEM for 3D-segmentation ........................................................................... 59 3.6 Artefact Mitigation Strategies ..................................................................................... 61 3.7 Summary ........................................................................................................................ 63 Chapter 4 Microstructural Analysis and Tortuosity Characterisation of Calendered NMC811 Electrodes 65 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals ........................... 66 4.1.1 Calendering and Microstructural Effects on Transport ................................. 67 4.1.2 Tortuosity: Definitions, Measurement, and Empirical Relations .................. 69 Table of Contents iii 4.1.3 Characterisation Techniques: X-ray CT versus FIB-SEM ............................... 71 4.1.4 Chapter Aims and Structure............................................................................... 72 4.2 Methodology for Microstructure and Transport Analysis ..................................... 73 4.2.1 Electrode fabrication ............................................................................................ 74 4.2.2 Electrochemical cell assembly and protocols ................................................... 75 4.2.3 Segmentation and 3D Reconstruction of FIB-SEM Images ............................ 76 4.2.4 Tortuosity analysis (Dragonfly) ......................................................................... 78 4.2.5 Tortuosity analysis (Taufactor) .......................................................................... 81 4.2.6 Tortuosity analysis (EIS Symmetrical Cell) ...................................................... 84 4.2.7 Statistical analysis (EIS Symmetrical Cell) ....................................................... 87 4.3 Results and Discussion ................................................................................................ 87 4.3.1 Microstructure comparison between uncalendered and calendered single and polycrystalline NMC811 ............................................................................................. 89 4.3.2 Tortuosity Analysis: Dense vs Sparse Graphs ................................................. 93 4.3.3 Tortuosity Analysis: Throat weighted graphs ................................................. 95 4.3.4 Comparison of Taufactor and Dragonfly ......................................................... 96 4.3.5 Symmetrical EIS Analysis and the Bruggeman Relationship ........................ 99 4.3.6 Electrochemical Implications ........................................................................... 102 4.3.7 Tortuosity Anisotropy and Transport Mechanisms in SC and PC NMC811 Electrodes ............................................................................................................................ 104 4.3.8 Electrode Optimisation ..................................................................................... 113 4.4 Conclusion ................................................................................................................... 113 Chapter 5 Probing Charging Protocol-Induced Redox Transformations in NMC811 Using EELS 115 5.1 Introduction to Charging Protocols and EELS Analysis for NMC811 Degradation ....................................................................................................................................... 116 iv Table of Contents 5.2 Electrochemical Analysis ........................................................................................... 117 5.2.1 Charging Protocols ............................................................................................ 118 5.2.2 Electrochemical Impedance Spectroscopy (EIS) ............................................ 119 5.2.3 Electrochemical Results .................................................................................... 119 5.3 Quantitative EELS Mapping of Redox States in NMC811 .................................... 120 5.3.1 Redox Chemistry in Discharged NMC811 Cathode Material ..................... 120 5.3.2 Experimental Limitations in Oxidation State Discrimination ..................... 121 5.3.3 Pre-processing: Thickness Normalisation and Background Removal ....... 124 5.3.4 Gaussian Fitting ................................................................................................. 128 5.3.5 Quantitative Metrics for Oxidation State Determination ............................. 131 5.4 EELS Results ................................................................................................................ 134 5.4.1 Spatial Evolution of Transition Metal Redox and Oxygen Chemistry ....... 134 5.4.2 Protocol-Dependent Degradation Gradients ................................................. 135 5.4.3 Peak Centre Position Analysis ......................................................................... 140 5.4.4 Direct Structural Visualisation of Phase Transformations ........................... 143 5.5 Integrating Spectroscopic and Electrochemical Insights ...................................... 146 5.6 Conclusion ................................................................................................................... 147 Chapter 6 Classification of Nickel Oxidation States in NMC811 via EELS and Convolutional Neural Networks 149 6.1 Introduction: Challenges in Direct Oxidation State Classification ...................... 150 6.2 Challenges in Training Data Preparation for Machine Learning ........................ 152 6.2.1 Limitations of the L3/L2 ratio approach ........................................................... 152 6.2.2 Modification to feature selection ..................................................................... 154 6.2.3 L3 Peak Area Analysis as an Alternative Approach ...................................... 156 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification .. ....................................................................................................................................... 158 Table of Contents v 6.3.1 Data Preparation for Machine Learning ......................................................... 159 6.3.2 Convolutional Neural Network Architecture ................................................ 160 6.3.3 Training and Validation Methodology ........................................................... 163 6.4 Classification Results and Discussion...................................................................... 164 6.4.1 Accuracy of model ............................................................................................. 164 6.4.2 Spatial Mapping of Oxidation States in NMC811 ......................................... 167 6.4.3 Limitations and Future Opportunities for CNN-Based Classification ....... 172 6.5 Relating EELS-CNN Findings to Other Spectroscopic Techniques ..................... 173 6.5.1 Similarities Between EELS and XAS for Oxidation State Analysis ............. 173 6.5.2 Adapting CNN Classification Across EELS and XAS: Complementarity and Compatibility ..................................................................................................................... 174 6.5.3 Prospects for Unifying EELS and XAS Through CNN Classification ........ 174 6.6 Conclusion ................................................................................................................... 175 Chapter 7 Conclusions and Future Works 177 7.1 Summary of Methodological Contributions ........................................................... 177 7.1.1 Three-Dimensional Microstructural Characterisation .................................. 177 7.1.2 EELS Analysis and Multi-Gaussian Fitting .................................................... 178 7.1.3 Machine Learning Classification of Oxidation States ................................... 179 7.2 Recommendations for Microstructural Characterisation ..................................... 179 7.3 Recommendations for Spectroscopic Analysis ....................................................... 180 7.4 Open-Source Software and Reproducibility ........................................................... 181 7.5 Final Reflection ........................................................................................................... 181 References 183 List of Figures 206 List of Tables 220 Appendix 222 Introduction 1 Chapter 1 Introduction Achieving net-zero emissions by 2050 is one of the defining global challenges of our time. A central pillar of this transition is the decarbonisation of the transport sector, primarily through the widespread adoption of electric vehicles (EVs). According to the International Energy Agency (IEA),1 EVs must account for 65% of global car sales by 2030, up from just 18% in 2023, to stay aligned with climate goals. This transition is already underway, with global electric car sales growing dramatically from 0.1 million units in 2012 to over 16 million units in 2024 (Figure 1.1), led predominantly by China, Europe, and the United States. Despite this impressive trajectory, further acceleration of EV adoption is needed, placing unprecedented demands on energy storage technologies, particularly lithium-ion batteries (LIBs), which underpin modern EV performance.2 Figure 1.1: Global electric car sales by region from 2012 to 2024, showing exponential growth from 0.1 million in 2012 to over 16 million in 2024. For the year 2024, China leads with approximately 10 million sales, followed by Europe (3.4 million), the United States (1.7 million), and the rest of the world (1.4 million). Data source: IEA 3 2 Introduction While fossil fuels have dominated road transport for over a century, the concept of electric vehicles dates back to the late 19th century. Early EVs were powered by lead-acid batteries with limited energy densities of around 40 - 60 Wh kg-1,4 which severely constrained driving range. As expectations for speed and reliability increased, “range anxiety”, which is the fear of running out of power before reaching a destination, emerged as a major obstacle to widespread EV adoption. The late 20th century saw renewed interest in EVs, alongside the development of improved battery chemistries. A notable example was General Motors’ EV1, launched in 1997, which initially relied on lead-acid batteries and offered a range of just 74 miles.5 The later introduction of nickel-metal hydride batteries extended the range to 150 miles,5 but these early EVs still lagged behind internal combustion engine vehicles in performance, range, and cost. It was only with the commercialisation of LIBs that EVs began to compete seriously with conventional vehicles.6 Modern EVs now rely on LIBs designed for high energy and power density. Among various cathode chemistries, nickel-rich layered oxides such as LiNi1-x-yCoxMnyO2 (NMC) have become especially prominent due to their ability to deliver high cell-level energy densities of around >150 Wh/kg.7 This has enabled ranges of up to 400 miles per charge, effectively resolving the limitations that once defined early EVs.8–10 As the battery industry matures, the focus has shifted beyond energy density to include cost, lifespan, and material sustainability.2 LIB prices have declined significantly in recent years, driven by advances in manufacturing, increased scale, and competitive supply chains.1 Figure 1.2 shows the evolution of raw material costs, including lithium carbonate, cobalt sulfate, and nickel sulfate, alongside average battery pack prices over the past decade. Although material costs have fluctuated, particularly between 2021 and 2022, battery pack prices have continued to fall. According to Goldman Sachs Research, these prices are expected to drop to continuously over the next few years, bringing EVs closer to cost parity with internal combustion engine vehicles on a total cost-of-ownership basis.11 Nonetheless, affordability alone is not sufficient to ensure sustainability. The sourcing of raw materials, especially cobalt, raises significant ethical and environmental concerns.13 Early NMC chemistries such as NMC111 relied heavily on cobalt, but the industry has since moved toward cobalt-lean, nickel-rich variants like NMC811. These materials offer higher energy density and reduced reliance on cobalt but introduce new challenges related to chemical and structural stability.7 Introduction 3 Figure 1.2: Price trends of selected battery materials and lithium-ion battery packs from 2015 to 2024. The index tracks prices of key raw materials including cobalt sulphate, lithium carbonate, nickel sulphate, manganese flake, and phosphoric acid. Data source: IEA12 NMC811, used in commercial EVs such as the BMW ix3 and Volkswagen ID.3,14,15 exemplifies these trade-offs. While offering superior energy performance, its high nickel content increases reactivity, especially at high states of charge, leading to degradation phenomena such as surface reconstruction, microcrack formation, and active lithium loss.16–24 These degradation processes span multiple length scales, from atomic-level redox changes to microstructural evolution, and are difficult to fully characterise without advanced analytical tools. To understand these complex failure modes, researchers increasingly rely on high-resolution techniques such as focused ion beam scanning electron microscopy (FIB-SEM)25and electron energy-loss spectroscopy (EELS).26,27 These tools can generate detailed spatial and spectral data, but their interpretation is often labour-intensive, time-consuming, and varies between analysts. This creates a bottleneck in data analysis, particularly as the demand for EV batteries is projected to grow from 340 GWh in 2022 to over 3,500 GWh by 2030.12 The overarching aim of this thesis is to tackle the analytics bottleneck that increasingly limits what can be learned from modern electron microscopy and spectroscopy of battery materials. Using NMC811 cathodes as a technologically relevant case study, the work develops and validates robust, reproducible data-analysis methodologies that allow complex, information- rich datasets to be interpreted more consistently and at higher throughput. A key motivation throughout is to move beyond treating microscopy as primarily qualitative, and instead extract quantitative descriptors that can be compared reliably across samples, processing 4 Introduction conditions, and cycling protocols. To achieve this, the thesis combines physics-informed analysis with rigorous statistical fitting and machine learning, with the specific goal of reducing analyst-dependent interpretation while improving reproducibility in both microstructural and spectroscopic characterisation. At the electrode scale, this thesis establishes a FIB-SEM tomography workflow that links three- dimensional microstructure directly to transport behaviour. Segmented reconstructions are used not only for visualisation, but as the basis for quantitative analysis of connectivity and directional transport. By combining graph-based connectivity metrics with multiple tortuosity calculation approaches and finite-difference simulations, and benchmarking these insights against electrochemical impedance spectroscopy, the workflow provides a systematic route to quantify anisotropy and to relate manufacturing variables, such as calendering, to changes in ion transport pathways across different electrode architectures. At the nanoscale, the thesis develops an EELS analysis pipeline designed to cope with the practical realities of battery datasets, where spectra are often acquired in high volume and under non-ideal conditions, with complex backgrounds and overlapping features. Novel background fitting strategies and a comprehensive multi-Gaussian fitting methodology are introduced to enable more reproducible quantification of transition-metal oxidation states and to probe changes in oxygen coordination across degradation interfaces. By prioritising robustness and scalability, the approach supports the systematic analysis of spectral datasets that would otherwise require extensive manual processing and subjective decision-making. Building on these foundations, the thesis then introduces a convolutional neural network framework for automated classification of nickel oxidation states directly from EELS spectra. This shifts the analysis away from expert-driven interpretation of subtle spectral differences and towards a high-throughput, repeatable route for mapping redox gradients. By reducing reliance on subjective judgement and enabling rapid processing of large spectral datasets, the framework supports more consistent comparison across experimental conditions and provides an approach that can keep pace with modern microscopy acquisition. Taken together, these contributions strengthen the ability of electron microscopy to deliver not only high-resolution images and spectra, but also statistically robust, performance- relevant metrics that can be used to interrogate degradation in Ni-rich cathodes. More broadly, the workflows developed here address a practical need in battery research: analytical tools that translate complex experimental datasets into quantitative insights that can inform materials design, manufacturing optimisation, and protocol development for longer-lasting NMC811 electrodes. To support reproducibility and reuse, all code developed in this work has been made openly available via GitHub. Introduction 5 Following the literature review in Chapter 2, which frames the evolution of cathode chemistries, the dominant degradation modes in Ni-rich layered oxides, and the analytical challenges posed by modern microscopy datasets, Chapter 3 describes the experimental and analytical foundation of this work. It outlines the electron microscopy and spectroscopy methods used throughout the thesis, with particular emphasis on FIB-SEM tomography and STEM-EELS, and details the specialist sample preparation and acquisition strategies needed to obtain reliable, comparable datasets from degradation-sensitive NMC811 specimens. Building on this platform, the thesis then progresses from electrode-scale transport to nanoscale chemistry in a connected analysis thread. Chapter 4 quantifies how manufacturing, specifically calendering, reshapes three-dimensional electrode microstructure and directional transport behaviour in both single-crystal and polycrystalline NMC811 architectures by integrating segmentation-led reconstruction with tortuosity calculations, graph-based connectivity metrics, finite-difference simulations, and electrochemical impedance spectroscopy. With these microstructural descriptors established, Chapters 5 and 6 focus on the chemical dimension of degradation: Chapter 5 develops and applies a robust EELS processing and multi-Gaussian fitting methodology to quantify redox evolution and oxygen coordination across degradation interfaces under different charging protocols, while Chapter 6 extends this approach into high-throughput automation via a convolutional neural network for nickel oxidation-state classification directly from EELS spectra. Finally, Chapter 7 synthesises the microstructural and spectroscopic findings to highlight the broader implications for electrode design and operation, and to outline how the analysis frameworks introduced here can enable more scalable, reproducible characterisation of Ni-rich cathode degradation moving forward. 6 2.1 Fundamentals of Lithium-ion Battery Operation Chapter 2 Cathode Materials for Lithium-ion Batteries: Evolution, Challenges, and Characterisation 2.1 Fundamentals of Lithium-ion Battery Operation Lithium-ion battery (LIB) operation relies on the controlled shuttling of lithium ions between two porous electrodes, an anode and a cathode, separated by an electronically insulating but ionically conductive separator filled with liquid electrolyte (Figure 2.1). This architecture creates two parallel transport networks: ionic transport through the electrolyte-filled pore phase, and electronic transport through the solid phase. Electrons flow through the external circuit to deliver power, while Li⁺ moves internally through the electrolyte within the separator and electrode pores. Each electrode comprises active material particles that store lithium, conductive additives that provide electronic percolation, and polymer binders that ensure mechanical integrity.28 Figure 2.1: Schematic representation of Li-ion battery operation. Red dots indicate Li-ion movement through the electrolyte, and green arrows show electron flow in the external circuit. The top panel illustrates charging, and the bottom shows discharging. Sizes are not to scale. Figure produced by Morzy.29 2.1 Fundamentals of Lithium-ion Battery Operation 7 During charging, an external power source applies a potential that drives electrons from the cathode through the external circuit to the anode. At the same time, lithium ions are deintercalated from the cathode and transported through the electrolyte, crossing the separator to intercalate into the anode. During discharge, both fluxes reverse: Li+ exits the anode and migrates back through the electrolyte towards the cathode, while electrons flow through the external circuit to power an external load. In both directions, the electrochemical reactions are only accessible if ions and electrons can reach reaction sites throughout the electrode thickness. 28 The overall cell reaction during discharge can be generalised as:20,30 𝐿𝑖𝑀𝑂2 + 𝐶6 𝑐ℎ𝑎𝑟𝑔𝑒 ⇌ 𝑑𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒 𝐿𝑖𝑥𝐶6 + 𝐿𝑖1−𝑥𝑀𝑂2 𝑨𝒏𝒐𝒅𝒆: 𝑥𝐿𝑖+ + 𝐶6 + 𝑥𝑒− 𝑐ℎ𝑎𝑟𝑔𝑒 ⇌ 𝑑𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒 𝐿𝑖𝑥𝐶6 𝑪𝒂𝒕𝒉𝒐𝒅𝒆: 𝐿𝑖𝑀𝑂2 𝑐ℎ𝑎𝑟𝑔𝑒 ⇌ 𝑑𝑖𝑠𝑐ℎ𝑎𝑟𝑔𝑒 𝐿𝑖1−𝑥𝑀𝑂2 + 𝑥𝐿𝑖+ + 𝑥𝑒− where M denotes transition metals such as Ni, Mn, and Co. While these reactions are thermodynamically reversible, practical battery performance and ultimately degradation is governed by the kinetics and transport limitations within porous composite electrodes. Understanding these transport fundamentals is essential for interpreting the degradation mechanisms discussed in Section 2.3. Lithium transport is also required within the active material. After charge transfer at the particle surface, lithium redistributes through the particle via solid-state diffusion. In layered NMC oxides, Li+ diffuses through channels between transition metal oxide sheets. At high current densities, transport limitations in either phase generate concentration gradients and non-uniform utilisation through the electrode thickness. Additionally, charge transfer kinetics at the electrode-electrolyte interface, often quantified by charge transfer resistance (𝑅𝐶𝑇), can become rate-limiting, especially when surface degradation layers form. 20 Electronic transport is controlled by connectivity of the solid phase. Many oxide cathodes have limited intrinsic electronic conductivity, so conductive carbon additives form a percolating network connecting particles to the current collector. Electronic transport is therefore sensitive to conductive additive distribution and to contact resistances at particle interfaces.28,31 8 2.1 Fundamentals of Lithium-ion Battery Operation The interdependence of these transport pathways shapes electrode design and processing strategies. For example, calendering compacts electrodes to improve particle contact and reduce electronic resistance, but simultaneously reduces pore volume and increases ionic tortuosity. Similarly, degradation mechanisms often affect both pathways: particle cracking disrupts electronic percolation while surface reconstruction layers impede ionic charge transfer (discussed in Section 2.3).32,33 Table 2.1: Summary of key battery performance metrics Metric Definition Unit Energy Density Energy stored per unit volume or mass of the battery. Crucial for compact, lightweight designs. 𝐸𝑛𝑒𝑟𝑔𝑦 𝐷𝑒𝑛𝑠𝑖𝑡𝑦 = ∫ 𝑉𝑐𝑒𝑙𝑙 × 𝐶𝑐𝑒𝑙𝑙 Wh L-1 or Wh kg-1 Cell Voltage (𝑉𝑐𝑒𝑙𝑙) The operating voltage of a single cell during discharge, dictated by electrode potential difference. V Capacity (𝐶𝑐𝑒𝑙𝑙) Total amount of charge a cell can store. Reflects how much lithium can be intercalated. Ah or mAh g -1 Specific Energy Energy per unit mass, important for EV driving range. 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝐸𝑛𝑒𝑟𝑔𝑦 = 𝐸𝑛𝑒𝑟𝑔𝑦 / 𝑀𝑎𝑠𝑠 Wh kg -1 Specific Power Rate of energy delivery per unit mass, affecting acceleration potential. 𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐 𝑃𝑜𝑤𝑒𝑟 = 𝑃𝑜𝑤𝑒𝑟 / 𝑀𝑎𝑠𝑠 W kg-1 Cycle Life Number of full charge-discharge cycles before performance drops below a threshold. Cycles 2.2 Evolution of Cathode Materials 9 These transport fundamentals motivate the characterisation approaches used later in this thesis. When ionic or electronic transport becomes limiting, local current density and state-of- charge can become spatially heterogeneous, which promotes non-uniform ageing. This motivates the multi-scale characterisation strategy employed in this thesis: three-dimensional microscopy (FIB-SEM) to quantify microstructure-dependent descriptors such as tortuosity and connectivity at the electrode scale, and spatially resolved spectroscopy (EELS) with robust quantitative analysis methods to map oxidation-state evolution and chemical heterogeneity at the nanoscale.34,35 To better understand how different cathode materials perform and compare, a summary of key LIB performance metrics is provided in Table 2.1. These metrics help to contextualise the strengths and limitations of each material and provide a framework for evaluating the trade- offs discussed in the next section. 2.2 Evolution of Cathode Materials Although all components contribute to functionality, the cathode remains the performance- limiting electrode due to its lower theoretical capacity (~274 mAh g-1 for LiCoO2 versus ~372 mAh g-1 for graphite), its contribution to 25 - 30 % of the total cell cost, and its role as the primary source of thermal instability and degradation challenges.36,37 Furthermore, cathode synthesis demands stricter control compared to other components. As such, improving cathode materials represents the most significant opportunity for advancing battery technologies, with particular efforts directed towards increasing energy density while reducing reliance on cobalt for cost and sustainability reasons. These developments are summarised in Figure 2.2. 10 2.2 Evolution of Cathode Materials Figure 2.2: Historical milestones of the Ni-rich cathode development. The abbreviations NMC consist of LiNi1-x-yMnxCoyO2 where references were taken from NMC11138, NMC42239, NMC532,40 NMC622,41 NMC811,22 NMC91.13 Reference used for LiNixMn1-xO242 and LiNixCo1- xO2 (LiNi0.75Co0.25O2).43 References used for LiCoO2.6,44 Figure produced by Wu et al.45 2.2.1 Early Cathode Materials LiCoO2 (LCO), developed by Goodenough, Whittingham, and Yoshino (joint 2019 Nobel Prize in Chemistry recipients) and commercialised by Sony in the early 1990s, represented the first successful commercial cathode.6 Despite its adoption, LCO suffers from structural degradation at high states of charge (Li1-xCoO2 where x > 0.5). This is when significant amounts of Li-ion have been extracted and the CoO2 framework can no longer maintain structural integrity, leading to the release of lattice oxygen. LCO also suffer from limited practical capacity (145 mAh g-1 vs theoretical 274 mAh g-1) and reliance on expensive cobalt with significant ethical mining concerns.46 The mid-1990s saw the emergence of alternative cathode chemistries. LiFePO4 (LFP) offered superior structural and thermal stability with excellent cycling life due to strong Fe-O covalent bonds. A drawback is that LFP has lower power density when compared to LCO due to its lower operating voltage, 3.3V and 3.9V respectively.47 Concurrently, LiMn2O4 (LMO) provided higher voltage (4 V vs. Li/Li+) leading to enhance power and energy density whilst at a lower cost due to the cheaper raw material of Mn. However, LMO suffers from capacity fading due to Mn dissolution and spinel phase 2.2 Evolution of Cathode Materials 11 transformation.47 While these materials improved certain attributes such as cost and thermal safety, they generally exhibited lower energy densities compared to layered oxides, driving continued efforts to optimise cathode materials. 2.2.2 Development of Mixed Transition Metal Oxide The limitations of single-metal oxides spurred the development of mixed-metal systems. Substituting cobalt with nickel produced LiNiO2 (LNO) which maintained the layered structure of LCO while offering higher capacity (~250 mAh g-1) and lower cost.38,48 However, LNO’s practical application was hindered by synthesis challenges, Ni3+ instability and significant cation mixing.49 This led to the development of LiNi1-yCoyO2, where Co3+ substituted the unstable Ni3+ to improve structural integrity and stability. Further refinement produced LiNi1-x-yCoxAlyO2 (NCA) through dual doping of Co and Al, providing high capacity (200 mAh g-1) with improved thermal stability, though still suffering from capacity fade at elevated temperatures and high currents.50 The late 1990s marked a significant breakthrough with LiNi1-x-yCoxMnyO2 (NMC) development.51 By the early 2000s, LiNi1/3Co1/3Mn1/3O2 (NMC111) dominated early EV batteries,22 offering a balanced composition where each element contributed specific properties to overall performance. 2.2.3 Emergence of Ni-rich NMC Cathodes The mid-2000s marked a crucial advancement with the introduction of Ni-rich layered cathode materials, particularly NMC811 (LiNi0.8Co0.1Mn0.1O2). This composition delivered high specific capacity (>200 mAh g-1), an elevated operating potential (>3.8 V vs. Li/Li⁺), and reduced cobalt content, aligning with both performance and sustainability goals.52 Within NMC811, each transition metal fulfils a distinct role. Nickel, comprising 80% of the transition metal layer, provides the primary redox activity via multiple oxidation states (Ni2+/Ni3+/Ni4+), underpinning its high capacity.53 However, the higher fraction of Ni4+ at elevated states of charge enhances material reactivity, promoting surface degradation. Manganese, primarily present as Mn4+, stabilises the layered structure and improves thermal stability.54 Elevated Mn concentration at the surface also suppresses interfacial side reactions, mitigating impedance growth.55 Despite its reduced proportion (10%), cobalt remains crucial for electronic and ionic conductivity enhancement and structural stabilisation, helping to inhibit detrimental Ni2+/Li+ disorder.56 Furthermore, cobalt promotes the formation of a Co3O4- 12 2.2 Evolution of Cathode Materials type spinel structure under thermal stress, which inhibits transformation to the electrochemically inactive rock-salt phase.57 Figure 2.3: (a) Comparison of Ni4+ concentration of various NCM cathode against different depth of charge. Copyright 2017, American Chemical Society.24 (b) Capacity retention with the increased concentration of Ni. Copyright 2018 Chemistry of Materials.58 The performance trade-offs associated with increasing Ni content are illustrated in Figure 2.3. As shown in Figure 2.3a, Ni4+ concentration rises with Ni-rich compositions, increasing reactivity and degradation risk. Capacity retention trends (Figure 2.3b) reveal declining stability with higher nickel content, particularly in fully substituted LiNiO2, highlighting the challenge of managing degradation in Ni-rich cathodes.45 Overall, the evolution of cathode materials has been driven by the need to balance energy density, stability, cost, and sustainability. While NMC811 represent a major achievement, offering high specific energy and reduced cobalt content, these advances come with inherent trade-offs. In particular, the high reactivity of nickel-rich surfaces and structural instability at elevated states of charge introduce challenges for long-term cycling performance, which constitutes as a barrier for widespread EV adoption. As such, understanding the complex degradation mechanisms in Ni-rich cathodes is essential for further improving their viability in next-generation batteries. The following section focuses on the degradation pathways of NMC811, examining how structural, chemical, and electrochemical changes compromise performance over time. 2.3 Degradation of NMC811 13 2.3 Degradation of NMC811 The degradation of NMC811 is a complex, multifaceted phenomenon comprising several interconnected mechanisms. Although these processes are typically investigated individually, due to experimental limitations that restrict probing to specific aspects, it is crucial to recognise that they operate concurrently and often accelerate one another. Under high-voltage operation, Ni-rich layered cathodes predominantly suffer from four major degradation pathways: surface reconstruction, gas evolution, bulk phase transformation, and microcrack formation16,20–23,58–60. The following sections examine these degradation mechanisms in detail, with particular emphasis on their manifestation in NMC811. 2.3.1 Bulk Phase Transformation at High Voltage Throughout this work, the state of charge (SOC) refers to the extent of delithiation, where SOC = 1 denotes the fully charged (fully delithiated) state and SOC = 0 denotes the fully discharged (fully lithiated) state. Higher SOC reflects greater lithium extraction from the cathode. During electrochemical cycling, NMC811’s layered crystal structure undergoes dynamic changes, expanding and contracting with lithium content, a phenomenon often described as “lattice breathing” 61. As Figure 2.4a shows, the 𝑎-axis lattice parameter decreases steadily from 2.87 Å at 3.6 V (SOC = 0) to 2.81 Å at 4.4 V (SOC = 1). Meanwhile, the 𝑐-axis initially expands from 14.2 Å to 14.5 Å up to ~70% SOC, driven by increased electrostatic repulsion between oxygen layers as lithium ions are removed. Beyond ~70% SOC (around 4.1 V), the c-axis collapses abruptly to ~14.0 Å, a phenomenon known as lattice collapse53,62. The differential capacity plot in Figure 2.4b further illustrates this behaviour. During charging, NMC811 undergoes a series of phase transitions: hexagonal (H1) → monoclinic (M) → hexagonal (H2 and H3). 61,63,64 Above 4.11 V, corresponding to the H2-H3 transition, significant cell volume contraction occurs alongside oxidation of lattice oxygen, often leading to secondary particle cracking. This H2-H3 region is structurally unstable and undermines the reversibility of lithium storage. 14 2.3 Degradation of NMC811 Figure 2.4: (a) Evolution of average NMC811 unit cell volume and lattice parameters (𝑎 and 𝑐) during delithiation measured using X-ray diffraction29,53 (b) dQ/dV profiles of NMC811 between 2.9 and 4.7 V vs. Li/Li+.16 Operando X-ray diffraction and X-ray absorption spectroscopy studies by Kondrakov et al.62 revealed that the transition metal slab thickness decreases steadily by ~16 %, while the lithium slab initially expands to ~120 % of its original height before contracting to ~96 % as SOC increases. These slab changes reflect evolving electrostatic interactions: at low SOC, lithium ions screen the negatively charged oxygen lattice, but as delithiation proceeds, weakened screening enhances oxygen-oxygen repulsion, driving c-axis expansion. At high SOC, charge transfer from oxygen 2p to nickel 3d eg orbitals reduces oxygen electron density, triggering 𝑐- axis collapse. This lattice breathing has significant implications for ionic transport. Marker et al.53 showed that lithium mobility initially improves due to vacancy formation and expanded interlayer spacing but declines sharply after lattice collapse, impeding diffusion. Moreover, the anisotropic strain generated by these lattice changes contributes to mechanical degradation, including microcracking. 2.3.2 Microcracks Mechanical degradation through microcrack formation plays a major role in NMC811 performance decline. These cracks emerge during manufacturing and intensify during cycling due to volumetric changes that generate stresses exceeding material tolerance.65 Microcracks manifest primarily in two forms: 2.3 Degradation of NMC811 15 Intergranular cracking occurs along grain boundaries between primary particles and represents the dominant mechanical degradation mechanism66. In Ni-rich cathodes, secondary particles typically form through agglomeration of smaller primary grains, creating numerous vulnerable boundaries. Charge/discharge-induced anisotropic volume changes generate strains at these boundaries,62,66,67 particularly during the H2-H3 phase transition at approximately 4.1 - 4.2 V, where materials with nickel contents of 80% and above are especially vulnerable.21 Intragranular cracking develops within individual primary particles.68 These fractures result from stress concentration and inhomogeneous delithiation. The mechanical stability of NMC811 relies heavily on Li-O bonds: tensile stress primarily compromises TM-O bonds in- plane, but Li-O bonds dominate out-of-plane dissociation.69 Compression shortens Li-O bonds, reducing mechanical stability in all directions.68–70 Inhomogeneous lithium distribution exacerbates stress concentrations, accelerating crack formation.71 Microcracks expose fresh surfaces to electrolyte, thickening the cathode- electrolyte interphase (CEI), promoting rock-salt formation, increasing impedance, and disconnecting active particles, ultimately degrading cycling and thermal stability.23 2.3.3 Ni/Li site exchange Before cycling, in its pristine condition, NMC811 adopts a layered α-NaFeO₂-type structure (space group 𝑅3̅𝑚), where LiMO₂ (M = Ni, Mn, Co) forms alternating layers of edge-sharing LiO₆ and MO₆ octahedra (Figure 2.5a). Lithium ions occupy 3a sites, transition metals occupy 3b sites, and oxygen ions occupy 6c sites.72,73 Due to the similar ionic radii of Li+ (0.76 Å) and Ni2+ (0.69 Å ),74 Ni²⁺ can migrate into lithium sites, resulting in cation mixing or Ni/Li site exchange.75 This disorder reduces lithium-ion mobility by blocking two-dimensional diffusion pathways and limits the amount of lithium that can be reversibly inserted during discharge. Cation mixing tends to worsen at high states of charge, where structural destabilisation promotes Ni migration (Figure 2.5c,d). Additionally, cation disorder can form during synthesis, particularly in oxygen-deficient environments that lower the migration barrier.75 To suppress disorder, calcination should be performed under oxygen-rich conditions with carefully optimised lithium precursor ratios and sintering profiles. 16 2.3 Degradation of NMC811 Figure 2.5: (a) Ordered 𝑅3̅𝑚 structure. (b) Cation mixing phase with 𝐹𝑚3̅𝑚 structure. (c) 𝑅3̅𝑚 structure with Li vacancies at highly charged state. (d) Cation mixed phase with partial TM ions in Li layer. Liu et al..76 Ni/Li disorder is often more pronounced near particle surfaces, where lithium depletion and structural defects accelerate Ni migration.76 Microcracking exacerbates this process by exposing fresh surfaces. Over cycling, disorder accumulates, raising charge-transfer resistance and contributing to capacity fade. In severe cases, extensive Ni occupation of Li sites leads to a transformation of the layered 𝑅3̅𝑚 structure into the cubic 𝐹𝑚3̅𝑚 rock-salt phase Figure 2.5b, forming electrochemically inactive "dead zones" at the particle surface, often referred to as reduced surface layers or rock-salt layers. These regions permanently block lithium-ion transport, leading to irreversible capacity loss. Thus, Ni/Li site exchange is both a cause and consequence of structural degradation in NMC811, closely tied to synthesis conditions, cycling history, and mechanical damage.76 2.3.4 Surface Reconstruction While the bulk of NMC811 governs lithium intercalation, degradation at the electrod- electrolyte interface, commonly referred to as surface reconstruction, has emerged as a critical 2.3 Degradation of NMC811 17 contributor to long-term performance decline. At high states of charge (SOC > 75%), Ni-rich layered oxides undergo surface structural transformations from the pristine layered 𝑅3̅𝑚 phase to spinel-like and ultimately to rock-salt (𝐹𝑚3̅𝑚) phases, accompanied by lattice oxygen loss and transition metal (TM) migration. This results in the formation of an electrochemically distinct surface layer, often termed the reduced surface layer (RSL).77,78 These reconstructed layers impede lithium diffusion and increase charge-transfer resistance. When sufficiently thick, the RSL isolates the underlying active material, limiting the accessible SOC range.79 As illustrated in Figure 2.6, a thick rock-salt layer constrains 𝑐-axis contraction, leading to electrochemical fatigue that confines the SOC window to 0 - 75 %. Surface reconstruction has been extensively studied since 2013-2014.59,77,80 Jung et al.59 observed spinel- and rock-salt-like layers on NMC532 surfaces using HR-TEM, with thicker layers forming at higher cut-off voltages. Lin et al77 corroborated this via STEM-EELS, showing that even mere electrolyte exposure initiates TM reduction and RSL formation. With increasing UCV (4.3 - 4.7 V), the layer thickened, particularly on facets exposing lithium channels to the electrolyte. Electrochemical consequences of surface reconstruction include increased impedance and capacity fade, especially at high cut-off voltages. 77 The evolution of insulating layers at the surface slows lithium kinetics, raising resistance. Figure 2.6: The structural evolution during delithiation of the (e) active phase and (f) the fatigued phase. In the illustrations, lithium atoms are represented by green circles, transition metal atoms by blue circles, and oxygen atoms by red circles. Figure produced by Kleiner et al..79 18 2.3 Degradation of NMC811 Zhang et al.81 showed that RSLs only develop at particle surfaces exposed to electrolyte, while buried interfaces remain unaffected (Figure 2.7), underscoring the necessity of electrolyte contact for surface reduction. Notably, RSL formation is self-limiting, constrained by the slow kinetics of TM migration and oxygen diffusion.82,83 Crucially, the loss of lattice oxygen not only triggers transition metal migration but can also result in the evolution of molecular oxygen gas. This oxygen release is particularly dangerous, as it raises internal cell pressure, increases the risk of electrolyte decomposition, and can contribute to thermal runaway under abusive conditions. Oxygen release is therefore not only a structural issue but also a critical safety concern for Ni-rich cathodes. Figure 2.7: (a) HAADF-STEM images with the atomic resolution for detection of the surface of pristine NMC samples. The scale bars are 5 nm in left image and 1 nm in region 1 and 2, respectively. (b) Scheme of atomic configuration for bulk and surface in polycrystalline NMC particle. Figure produced by Zhang et al..81 2.3.5 Surface-bulk modulation Finally, it is essential to consider the interplay between surface and bulk degradation. Some studies argue that surface chemical instability is the dominant cause of capacity loss in Ni-rich NMCs.78,84 The formation of sluggish surface phases, such as thick cathode-electrolyte interphase (CEI) layers and reconstructed rock-salt-like surfaces, combined with subsequent lattice changes in the bulk, has been proposed as the origin of fatigue behaviour in these materials.60 Others suggest that the dominant degradation mechanism varies with nickel content.58 When the Ni content is above 80%, phase transitions at high states of charge (SOC) induce residual stresses that generate microcracks within the bulk, which eventually propagate to the surface. In contrast, when the Ni content is below 80%, surface degradation appears to be the primary driver of capacity fading. 2.3 Degradation of NMC811 19 Morphological defects such as cracks and surface heterogeneity contribute to impedance heterogeneity and uneven charge distribution.85 These defects promote non-uniform degradation, disrupt lithium and electron transport pathways, and lower sub-particle utilisation. It is therefore essential to recognise that surface chemistry and bulk microstructure are not independent; rather, they engage in a complex, dynamic interplay that governs the degradation behaviour of Ni-rich NCMs. In addition to electrochemical cycling, defects in cathode materials can also form during the synthesis process. During calcination, two competing reactions occur: the transition metal (TM) hydroxide precursor tends to decompose on its own, while at the same time, lithium ions react with the surface of the TM precursor to form a layered oxide structure.86 The balance between these reactions depends on the thermal reactivity of the TM precursor with different lithium sources. Generally, a layered oxide phase forms at the interface, while the TM precursor core undergoes self-decomposition, producing spatially heterogeneous intermediates. This process leads to the development of structural defects at multiple scales, including intragranular nanopores within primary particles and intergranular voids between secondary particles, both of which undermine structural integrity and cycling stability. Figure 2.8: Illustration of the surface-to-bulk reaction coupling effect during battery process. The arrow presents the interaction between the surficial chemistry and the bulk microstructure. Figure produced by Li et al..85 In summary, the degradation of Ni-rich NMC811 cathodes stems from the complex and interconnected evolution of surface instability, bulk phase transitions, microcrack formation, and cation disorder. As shown in Figure 2.8, the degradation of Ni-rich NMC811 cathodes stems from a complex, dynamic interplay between surface and bulk mechanisms, often initiated during synthesis and evolving throughout cycling. 20 2.3 Degradation of NMC811 2.3.6 Modification Strategies and Single-Crystal versus Polycrystalline Morphologies The coupled degradation processes described in Sections 2.3.1 to 2.3.5 have driven extensive efforts to stabilise Ni-rich layered cathodes through material modification. Three main approaches exist to enhance the performance of nickel-rich layered cathode materials: surface coating, doping, and microstructure control. Surface coating involves depositing a thin protective layer on cathode particle surfaces to improve stability and prevent unwanted side reactions with the electrolyte; common coating materials include metal oxides, phosphates, and fluorides.87–89 Doping introduces small amounts of foreign ions into the crystal structure to modify electronic and ionic conductivity, stability, and kinetics.90–93 Microstructure control tailors the crystal size, morphology, and orientation of cathode particles to improve electrochemical kinetics and stability.94–97 Different morphologies of NCM particles, such as nanosheets, nanoplates, nanorods, and hollow spheres, exhibit distinct electrochemical behaviours due to variations in surface area and diffusion lengths.98 Among these modification strategies, the difference between single-crystal (SC) and polycrystalline (PC) particles has attracted significant research interest. The electrochemical performance of Ni-rich layered oxide (NCM) cathodes, including specific capacity, cycle life, and rate capability, is strongly dependent on crystal structure, particle size, and surface properties.99 NCM cathode materials consist of nanoscale layers of metal oxide separated by an interstitial Li+ ion layer, where crystal structure and composition significantly influence electrochemical performance. Additionally, crystal orientation and morphology play important roles in determining electrochemical behaviour.100 Polycrystalline (Figure 2.9, right) materials consist of numerous randomly oriented crystals with grain boundaries, whilst single-crystal materials are composed of perfect crystal lattices with few defects and almost no grain boundaries.101 Traditionally, NMC811 cathodes are polycrystalline, comprising spherical secondary particles composed of smaller individual primary particles. Whilst this hierarchical structure is cost-effective, scalable, and offers the lowest surface area-to-bulk ratio, it is mechanically fragile, especially during repeated charge- discharge cycles. Grain boundary misalignment in PC particles exacerbates this fragility, leading to cracking that exposes fresh surfaces to the electrolyte and promotes parasitic side reactions that reduce performance. In contrast, SC morphologies (Figure 2.9, left), characterised by larger, more uniformly sized primary particles, exhibit minimal grain boundary stress.102 This enhanced mechanical integrity translates to reduced particle cracking, improved Li-ion diffusion, and superior cycling performance under increasingly demanding conditions. Academically, SC particles also serve as a clean model system that enables precise investigation of fundamental mechanisms unobscured by grain boundary effects. 2.3 Degradation of NMC811 21 Figure 2.9: Comparison of single crystal (SC, left) and polycrystalline (PC, right) NMC811 cathodes after calendering. Green ticks and red crosses denote microstructural advantages and disadvantages, respectively. The definition of “single-crystal” (SC) morphology in Li-ion cathodes remains debated, and the term is used inconsistently across the literature. Mesnier and Manthiram103 therefore proposed an operational definition based on two criteria: particles should be largely deagglomerated (any agglomerates comprising fewer than five crystals) and have an average particle size > 1 μm, distinguishing them from nanoparticles that typically exhibit poorer tap density and higher surface reactivity.104 (003)-oriented SC-NMC cathodes demonstrate superior capacity retention and rate capability compared to randomly oriented polycrystalline (PC) NMC cathodes, which suffer from preferential capacity fading and sluggish kinetics.105 The large volume changes and cracking of PC-NMC particles during cycling cause capacity fading and structural degradation, limiting their long-term stability.105 The primary distinction between SC and PC morphologies lies in surface area evolution during electrochemical cycling. While pristine SC and PC particles show no clear surface area trend, being largely determined by synthesis method, consistent differences emerge during cycling, typically investigated using gas adsorption data and the Brunauer-Emmett-Teller (BET) equation.106 Kr physisorption studies demonstrated that SC-NMC811 cathodes showed only slight decreases in BET surface area after charging to 4.2 V versus Li⁺/Li,106 aligning with the volumetric contraction during delithiation107 and absence of mechanical fracturing. In contrast, PC-NMC811 BET surface area increased dramatically (from 0.2 to 1.4 m2 g−1) at the same voltage, indicating that particle cracking dominates the surface area increase. Grain boundaries in PC-NMC811 appear to act as nucleation points where mechanical stress propagates into cracks.108 22 2.4 Importance of Electron Microscopy in Battery Cathode Studies Although increased surface area may seem beneficial for ionic diffusion by offering shorter path lengths and greater electrolyte coverage, it introduces significant challenges. Reactive surface sites, prevalent in highly oxidized Ni-rich transition metal oxides,109 promote parasitic reactions that negatively impact long-term electrochemical performance.110 The primary distinction between SC and PC morphologies lies in the evolution of their respective surface areas during electrochemical cycling. Whilst the surface areas of pristine SC and PC particles do not exhibit a clear trend, being largely determined by synthesis method, consistent differences emerge in how these surface areas evolve during cycling. These trends are typically investigated using gas adsorption data and the Brunauer-Emmett-Teller (BET) equation. 106 SC-NMC811 effectively mitigates microcracking compared to PC-NMC811.111 SC cathodes consist of isolated 1-2 𝜇𝑚 primary particles without internal grain boundaries, reducing surface reactivity and enhancing mechanical resistance.112,113 Moiseev et al.111 and Raju et al.114 observed that SC-NMC811 maintained excellent structural integrity after cycling and calendering, whereas PC samples exhibited severe breakage. SC-NMC811 also retained higher capacity after 300 cycles and exhibited smaller increases in charge-transfer resistance (𝑅𝐶𝑇), indicating fewer microcracks and more stable ionic and electronic pathways. Although PC materials initially offer higher discharge capacities at low currents, the superior mechanical integrity of SC structures makes them a promising strategy for high-energy Ni- rich cathodes. Understanding these fundamental differences requires advanced characterisation techniques like electron microscopy to probe structural changes at multiple length scales and visualize the microstructural features and degradation mechanisms governing cathode performance. 2.4 Importance of Electron Microscopy in Battery Cathode Studies Given the complex and multiscale degradation phenomena in Ni-rich cathodes like NMC811, advanced characterisation techniques are essential for understanding structural and chemical evolution. High-resolution electron microscopy enables direct visualisation of atomic-scale defects, while spectroscopic methods reveal local chemical changes, such as transition metal oxidation states and oxygen loss. Three-dimensional tomography further provides insights into the electrode microstructure, including particle connectivity, pore network geometry, and tortuosity, bridging the gap between microstructural features and electrochemical performance. 2.4 Importance of Electron Microscopy in Battery Cathode Studies 23 2.4.1 High Resolution Imaging High-resolution transmission electron microscopy (HR-TEM) and scanning TEM (STEM) enable atomic-scale visualisation of structural changes in Ni-rich cathodes such as NMC811. These techniques allow direct observation of phase transformations, lattice distortions, and mechanical damage with sub-nanometre resolution. In particular, high-angle annular dark- field (HAADF) STEM imaging, based on atomic number (Z-contrast), offers composition- sensitive imaging that differentiates transition metal-rich regions from lithium-rich or oxygen- deficient domains. HR-TEM imaging has proven instrumental in revealing the progression of degradation phenomena, such as the transformation from the layered structure to spinel-like and rock-salt phases at particle surfaces.18,115 It also enables the identification of microcracks, dislocations, and other defects that emerge during electrochemical cycling. Surface layers often exhibit rock-salt character after extended cycling at high voltages, while cracks serve as preferential pathways for electrolyte penetration and accelerated phase transformation.19 However, while HR-TEM and STEM provide detailed structural information, their full interpretative power is realised when combined with spectroscopic techniques. By integrating structural imaging with chemical mapping, researchers can directly correlate morphological changes with redox processes and compositional variations, offering a comprehensive picture of cathode degradation mechanisms. 2.4.2 Spectroscopic Information: EELS Electron energy loss spectroscopy (EELS) was developed by James Hillier and R.F. Baker in the mid-1940s,116 but it remained relatively underutilised for several decades. It was not until the 1990s, with major improvements in microscope instrumentation and vacuum technology, that EELS gained widespread adoption in research laboratories. With aberration correction and monochromator, modern EELS achieves spatial resolutions of 1 Å and energy resolutions down to a few meV, enabling sub-nanometre chemical analysis.117,118 In parallel with hardware developments, advances in data processing software, such as Digital Micrograph119 and HyperSpy120 have further enhanced EELS capabilities. While proprietary tools like Oxide Wizard121 built as a plug in to Digital Micrograph facilitate transition metal white-line analysis, open-source platforms offer greater flexibility for large datasets and customised workflows. Alongside hardware developments, advances in data analysis have greatly enhanced EELS capabilities. Digital Micrograph119 became a key platform for data acquisition, visualisation, 24 2.4 Importance of Electron Microscopy in Battery Cathode Studies and processing, with plug-ins such as Oxide Wizard121 enabling systematic characterisation of transition metal white lines, including extraction of L3/L2 ratios, onset energies, and peak widths for oxidation state mapping. However, Oxide Wizard’s simplified linear background fitting and limited scalability constrain its application to large spectral datasets. To address these challenges, open-source platforms like HyperSpy120 emerged, offering a modular Python-based framework built on libraries such as NumPy122 and SciPy.123 HyperSpy facilitates customised, high-throughput EELS workflows and enables extensible methods for white line analysis, providing a flexible alternative to proprietary software. Figure 2.10: (a) High-resolution transmission electron microscopy (HRTEM) image showing a cycled NMC811 particle with a reconstructed rock-salt surface layer (~12 nm thick) overlying the original layered structure. The insets show corresponding fast Fourier transform (FFT) patterns identifying the rock-salt (top) and layered (bottom) phases. (b) HAADF-STEM image of the same region with coloured boxes indicating EELS line-scan positions from surface to bulk. (c) EELS spectra of the O K-edge, Mn L-edge, Co L-edge, and Ni L-edge collected along the scan direction, revealing chemical variations with depth. Figure produced by Li et al.. 27 In STEM-EELS mode, a focused probe rastered across a specimen collects an energy-loss spectrum at each pixel, generating hyperspectral datasets. These enable two-dimensional chemical maps, such as Ni, Co, Mn L-edges and the O K-edge, which reveal compositional heterogeneities and local chemical environments.124 EELS is particularly powerful for mapping degradation in NMC811. Shifts in the Ni L3 edge towards lower energy, increases in 2.4 Importance of Electron Microscopy in Battery Cathode Studies 25 L3/L2 intensity ratios, and decreases in the O K-edge pre-peak intensity indicate transition metal reduction and oxygen loss, key signatures of surface reconstruction. While HRTEM can reveal structural phase transitions linked to degradation, it requires exceptionally thin samples, precise zone-axis alignment, and significant operator expertise. In contrast, STEM-EELS provides a more accessible method for quantifying chemical and structural degradation gradients, enabling nanometre-scale mapping of oxidation states and elemental distributions without the stringent demands of high-resolution imaging. This capability is critical for elucidating degradation phenomena in Ni-rich cathodes. Figure 2.11: (a) HAADF-STEM image of a cycled NMC811 particle after 20 cycles between 2.0 - 4.8 V at C/10 rate. (b) Corresponding EELS spectra showing the Mn, Co, and Ni L2,3 edges and O K-edge from the surface towards the bulk. (c) Evolution of L3/L2 intensity ratios for Mn, Co, and Ni as a function of depth. Figure produced by Hwang et al..26 For example, Figure 2.10a shows a cycled NMC811 particle where a ~12 nm thick reconstructed surface layer overlies the original layered structure.27 HAADF-STEM imaging (Figure 2.10b) and corresponding EELS spectra (Figure 2.10c) reveal reduced oxygen pre- peaks and nickel valence shifts at the surface. Similarly, Hwang et al26 showed that after 20 L3 L2 L3 L2 26 2.4 Importance of Electron Microscopy in Battery Cathode Studies cycles to 4.8 V, surface Ni reduction extended ~34 nm into the bulk (Figure 2.11), varying between 10-60 nm across different particles of the same sample. Analysis of the Mn, Co, and Ni L2,3 edges, together with the O K-edge (Figure 2.11c), revealed that while nickel underwent significant reduction, manganese and cobalt oxidation states remained largely unchanged. While X-ray absorption spectroscopy (XAS) provides oxidation state and local structure information, its bulk-averaged nature limits the resolution of nanoscale heterogeneities. In contrast, EELS, with sub-nanometre spatial resolution, enables direct observation of localised surface degradation, transition metal migration, and oxygen loss at individual grains and interfaces.125 Despite its capabilities, EELS faces several challenges. Plural scattering, where electrons undergo multiple inelastic scattering events, can broaden spectral features and complicate interpretation, particularly in thicker specimens.126–129 In addition, inaccuracies in background subtraction and variations in L3/L2 area quantification can introduce uncertainties into oxidation state analysis.130 Careful data processing is essential to mitigate these effects, including consistent removal of the zero-loss peak tail, fitting Gaussian or multiple Gaussian functions to the L2 and L3 peaks to accurately separate overlapping features, and calculating L3/L2 ratios under standardised conditions. Although open-source platforms such as HyperSpy120 provide useful tools, their limited customisability for specific fitting requirements motivated the development of a bespoke analysis code in this work (see Chapter 4). The L3/L2 ratio is particularly sensitive to the integration range and background subtraction methods employed, making comparisons across different studies challenging if protocols differ. Furthermore, variations in energy resolution and microscope calibration between instruments can cause shifts in edge positions and peak intensities, introducing additional variability. With rigorous experimental protocols, reliable determination of transition metal oxidation states from EELS is achievable. Ideally, reference samples with known oxidation states should be measured under identical experimental and analysis conditions to establish robust calibration baselines. Emerging machine learning approaches further promise to improve EELS data interpretation. 2.4.3 Data Analysis: Machine Learning The application of machine learning (ML) to electron energy loss spectroscopy (EELS) has advanced rapidly since 2010, driven by the increasing complexity and size of modern spectral datasets. A single STEM-EELS spectrum image can contain thousands of individual spectra, each rich in fine structure and chemical information. Analysing such high-dimensional data manually is impractical and prone to subjective bias, prompting the need for scalable, automated approaches. ML offers powerful solutions to extract meaningful chemical and 2.4 Importance of Electron Microscopy in Battery Cathode Studies 27 structural information, perform denoising, and classify oxidation states with greater speed and reproducibility.125 The timeline of machine learning development for EELS, highlighting the shift from simple noise filtering towards fully automated classification and prediction, is illustrated in Figure 2.12. Figure 2.12: Timeline showing the evolution of machine learning applications in EELS analysis from 2010 onwards. Initial methods focused on principal component analysis (PCA) and basic noise reduction. Subsequently, clustering algorithms and matrix factorisation methods enabled unsupervised chemical mapping. Since 2020, the introduction of deep learning approaches such as convolutional neural networks (CNNs), autoencoders, and generative models has enabled more sophisticated analysis, including automated oxidation state classification and cross-technique data generalisation.125,131–139 Broadly, ML algorithms can be categorised into supervised and unsupervised methods, depending on whether the model learns from labelled or unlabelled data, respectively. Unsupervised learning methods aim to uncover intrinsic patterns within unlabelled datasets. Early applications of unsupervised ML in EELS included Principal Component Analysis (PCA), which became a standard tool for denoising spectrum images by isolating high- variance components while discarding noise.134–136 Beyond PCA, clustering algorithms such as hierarchical agglomerative clustering and k-means clustering were employed to segment chemical phases, enabling unsupervised chemical mapping without prior knowledge of 28 2.4 Importance of Electron Microscopy in Battery Cathode Studies composition.132,133 For example, clustering strategies have been used to distinguish shell and core regions in bimagnetic nanoparticles, demonstrating the utility of unsupervised learning for spatially resolving complex chemical environments.132 Supervised learning, by contrast, involves training models on labelled datasets where the correct output is known. Early supervised applications in EELS included Support Vector Machines (SVMs), which identified optimal hyperplanes to separate spectral features corresponding to different oxidation states. SVMs have been successfully applied to classify Mn and Fe oxidation states from EELS white-line analysis.140 More recently, random forest models have been developed, achieving high predictive performance. Gleason et al.,125 for instance, reported a coefficient of determination (R2) of 0.85 and a root mean square error of 0.24 for Cu oxidation state prediction from simulated datasets. While traditional machine learning approaches show considerable promise, they are often limited in their ability to capture the hierarchical and high-dimensional nature of spectral data. In response, deep learning frameworks, particularly convolutional neural networks (CNNs), have been introduced for EELS analysis. CNNs, which extract features hierarchically through stacked convolutional layers, are especially well suited to recognising subtle spectral variations and complex patterns. CNN-based models have achieved significant milestones in spectroscopic characterisation. For example, Ji et al. developed MnEdgeNet,141 a CNN capable of decomposing mixed oxidation states of Mn from XAS and EELS L2,3 edges without external calibration, achieving an accuracy of 85 %. The capacity of CNNs to autonomously learn chemically relevant features makes them particularly attractive for oxidation state classification tasks. Despite their potential, challenges remain. Most CNN models are chemistry-specific and struggle to generalise across different transition metal systems. Variations in acquisition conditions, spectral noise, and thickness effects also impact model performance. Addressing these limitations is particularly important for Ni-rich systems like NMC811, where accurate tracking of Ni oxidation states is critical to understanding degradation. Recent efforts focus on developing CNNs specifically tailored for Ni oxidation state classification, improving robustness to noise and minimising pre-processing requirements. These approaches represent promising steps toward high-throughput, reproducible characterisation of cathode degradation mechanisms. 2.4 Importance of Electron Microscopy in Battery Cathode Studies 29 2.4.4 3-Dimensional Tomography: FIB-SEM Ion transport in battery electrodes is strongly influenced by the three-dimensional microstructure, particularly the tortuosity of ion pathways relative to electrode thickness142,143. Ni-rich cathodes like NMC811 often exhibit anisotropic transport properties due to manufacturing steps such as calendering, which mechanically densifies the electrode. Focused ion beam-scanning electron microscopy (FIB-SEM) tomography enables nanometre-scale 3D reconstructions of electrode microstructures, offering higher resolution than conventional X- ray computed tomography (XCT). FIB-SEM enables quantification of particle size distributions, phase connectivity, surface area densities, and critically, tortuosity.144–146 Figure 2.13: Illustration of a tortuous path of length 𝛥𝑙 through a porous microstructure of thickness 𝛥𝑥, where the shortest tortuous path is used to calculate the tortuosity of the sample. (a) Geometric perspective without constrictions, (b) Geometric perspective with constriction, where lighter pink region represent less weightage compared to darker pink regions, which indicate more constricted areas Accurate segmentation into active material, carbon-binder domain (CBD), and pore phases is essential for reliable microstructural analysis. Early studies employed greyscale thresholding to differentiate phases, but is often insufficient due to curtaining artefacts and overlapping greyscale intensities.147 To address these challenges, Ender et al.148,149 introduced silicone resin infiltration to enhance phase contrast and minimise curtaining artefacts. Although effective in mitigating charging, this approach suffered from incomplete pore filling and residual greyscale overlap between the resin and CBD, leading to segmentation uncertainties. More recently, supervised machine learning, particularly U-Net architectures implemented in platforms like Dragonfly,150 has enabled robust voxel classification even in challenging datasets. Following segmentation, geometric tortuosity151 (𝜏𝑔𝑒𝑜) is defined as the ratio of the shortest transport path length (𝛥𝑙 ) to the electrode thickness (𝛥𝑥 ), as illustrated in Figure 2.13a Constrictions and dead-ends increase effective path length, elevating tortuosity, illustrated in 30 2.4 Importance of Electron Microscopy in Battery Cathode Studies Figure 2.13b. Advanced computational tools such as Dragonfly150 and Avizo152 have enabled the detailed reconstruction of pore and active material, facilitating identification of bottlenecks and quantification of microstructural constraints. Beyond geometric analysis, physics-based simulations, such as TauFactor,153 solve Laplace’s equation on voxelised datasets to compute effective tortuosity. In this approach, segmented FIB-SEM datasets are converted into voxelised 3D stacks, and diffusion simulations are performed to calculate transport properties across each phase. TauFactor has been validated against other simulation platforms such as Avizo’s XLab Thermo package, demonstrating good agreement with experimental and numerical results.154 Figure 2.14: Flowchart illustrating methods for measuring tortuosity: Geometric approaches using 3D imaging for direct analysis, physics-based methods for employing numerical simulations and electrochemical techniques like diffusion and symmetrical cell tests. Feedback loops and statistical analyses link microstructure to macroscopic transport properties for material optimisation. 2.5 Summary 31 Complementary to imaging and simulation-based methods, experimental approaches such as diffusion method155,156 and the symmetrical cell method 157,158 have been widely used to measure bulk-effective tortuosity. However, they offer only spatially averaged values and cannot resolve local anisotropy or heterogeneity. Despite significant progress, no prior study has systematically combined graph-based connectivity analysis, TauFactor diffusion simulations, and electrochemical impedance spectroscopy to cross-validate tortuosity measurements from FIB-SEM reconstructions of NMC811 electrodes. This integrated approach bridges microstructural features with bulk transport properties, allowing investigation of anisotropic transport. A schematic overview of the methodology is shown in Figure 2.14, with its application discussed in Chapter 4. 2.5 Summary This chapter reviewed the evolution of lithium-ion battery cathode materials from early LCO to Ni-rich layered oxides such as NMC811. Fundamental operating principles were outlined, including the transport processes that govern electrode performance: ionic transport through electrolyte-filled pore networks and electronic transport through the solid-phase active material and carbon-binder domain. These processes, often characterised using ionic and electronic tortuosity, influence rate capability and are sensitive to electrode microstructure and manufacturing conditions. The review has examined the commercial drivers for high-nickel content cathodes, emphasising the benefits of increased energy density alongside the associated challenges of structural and chemical instability. Degradation mechanisms in NMC811, including surface reconstruction, bulk phase transitions, microcrack formation, and cation disorder, were examined in detail. The distinction between single-crystal and polycrystalline morphologies was discussed, highlighting how grain boundaries in polycrystalline particles provide pathways for electrolyte infiltration and crack propagation, while single-crystal architectures may offer improved mechanical integrity. Advanced characterisation techniques play a critical role in understanding these degradation pathways. Electron microscopy provides direct visualisation of structural evolution at atomic resolution, while EELS enables mapping of transition metal oxidation states and oxygen participation in redox processes. Three-dimensional tomography bridges microstructural features with transport properties, allowing quantification of tortuosity. Different measurement approaches, geometric analysis, physics-based simulations, and electrochemical methods, each offer distinct advantages, motivating the use of complementary techniques. 32 2.5 Summary This thesis addresses gaps in the current understanding through three experimental chapters: Chapter 4 employs FIB-SEM tomography to investigate microstructural evolution of NMC811 electrodes under different calendering conditions. By integrating graph-based analysis, finite- difference simulations, and electrochemical impedance spectroscopy, quantitative relationships between manufacturing parameters, microstructure, and transport behaviour are established. The chapter examines how single-crystal and polycrystalline morphologies respond differently to calendering, with implications for tortuosity anisotropy and electrochemical performance. Chapter 5 investigates the influence of charging protocols on redox evolution through spatially resolved EELS analysis. A multi-Gaussian fitting methodology enables consistent extraction of L3/L2 white-line ratios and peak positions from spectra acquired on non- monochromated microscopes. Applied to NMC811 cathodes cycled under different protocols, this approach reveals distinct surface-to-bulk redox gradients, demonstrating how prolonged high-voltage exposure is associated with more extensive surface reconstruction. Chapter 6 introduces a convolutional neural network for classification of nickel oxidation states from EELS spectra, addressing limitations of conventional analysis in distinguishing closely spaced valence states such as Ni2+ and Ni3+. This framework enables high-throughput analysis of redox gradients and reduces reliance on subjective spectral interpretation. Together, these chapters establish a set of electron microscopy analysis workflows spanning atomic-resolution spectroscopy to electrode-level tomography, providing validated tools for quantifying degradation in Ni-rich cathode materials. Overall, the central focus of this thesis is the development and validation of electron microscopy analysis methods for battery materials, spanning (i) 3D microstructure and transport quantification from FIB-SEM tomography and (ii) robust, higher-throughput oxidation-state mapping from EELS using reproducible fitting and machine learning. Chapter 3 33 Chapter 3 Experimental Methods This chapter outlines the experimental techniques employed to investigate the structural and electronic evolution of NMC811 cathode materials during electrochemical cycling. A multi-scale analytical approach was developed, combining advanced electron microscopy and spectroscopy methods to correlate local electronic structure changes with morphological degradation. Electron Energy Loss Spectroscopy (EELS) served as the primary analytical technique, enabling direct probing of transition metal oxidation states and oxygen participation in the charge compensation mechanism. This required development of specialised sample preparation methods using Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) to produce ultrathin specimens that minimise plural scattering effects. Additionally, extensive FIB-SEM tomography was performed to create three-dimensional reconstructions of the electrode microstructure, enabling quantitative analysis of tortuosity for both electron and ion transport pathways-critical parameters governing electrochemical performance and degradation mechanisms. The theoretical principles, instrumentation, and optimised experimental protocols for each technique are described, with particular emphasis on the methodological innovations developed to overcome challenges specific to NMC811 materials. The lamellae prepared using these methodologies are subsequently analysed with TEM techniques described in this chapter, while the advanced processing of EELS data is presented in Section 5.3 for the Gaussian fitting method and Section 6.3 for the machine learning approach to identify transition metal oxidation state. Similarly, the post-processing techniques applied to 3D segmentation images from FIB-SEM are detailed in Section 4.2.3. 34 3.1 Electron Microscopy 3.1 Electron Microscopy Electron microscopy techniques provide an ideal platform for investigating the complex degradation mechanisms in battery cathode materials, including NMC811, as they offer capabilities ranging from nanoscale morphological characterisation to atomic-resolution imaging and electronic structure analysis. These techniques are particularly powerful for battery materials research as they enable direct correlation between electrochemical cycling, microstructural evolution, and changes in electronic structure relationships that are fundamental to understanding capacity fade mechanisms. Electron microscopes range from compact bench-top scanning electron microscopes (SEM) to larger, more sophisticated transmission electron microscopes (TEM), which are often equipped with scanning transmission electron microscopy (STEM) capabilities. Each of these instruments can be outfitted with detectors to capture signals from scattered electrons, emitted X-rays, or energy losses from inelastic interactions. Recent advancements such as aberration correctors for atomic-resolution imaging and direct electron detectors for enhanced signal sensitivity have significantly extended the capabilities of electron microscopy.159 These developments make high-energy electron microscopy particularly powerful for investigating the nanoscale structure and chemistry of advanced battery materials, including NMC811. The working principle of all electron microscopy techniques centers around the interaction between incident electron beam and specimen, specifically exploiting elastic and inelastic scattering to generate signals (Figure 3.1), which, when collected by appropriate detectors, provide both imaging contrast and rich analytical information. 3.1.1 Electron Beam Interaction: Elastic vs Inelastic Scattering The interaction of electrons with matter can be described through the principle of wave- particle duality, where electrons exhibit both particle-like and wave-like behaviour.159 This dual nature provides complementary frameworks for understanding electron scattering phenomena, which can be classified in several ways. From an energy perspective, elastic and inelastic scattering describe interactions with negligible or measurable energy loss, respectively. From a wave perspective, coherent and incoherent scattering refer to whether electron waves remain in phase or become out of phase after interaction. Spatially, forward scattering occurs at angles less than 90° relative to the incident beam and is more common in thin specimens, while backscattering occurs at angles greater than 90° and happens in both thin and bulk samples. The different type of scattering phenomena are illustrated in Figure 3.1. 3.1 Electron Microscopy 35 Figure 3.1: Schematic representation of the electron beam interaction volume in a specimen, showing the various signals generated and their respective interaction depths160 Elastic scattering primarily results from Coulomb interactions between probe electrons and the electrostatic potential of atomic nuclei, commonly described as Rutherford scattering161. Due to the significant mass difference between electrons and nuclei, the electron trajectory changes while energy loss remains negligible. Elastic scattering is generally coherent in nature and predominantly forward-peaked, typically occurring at deflection angles less than 10°. Under parallel beam illumination, diffraction from crystalline specimens produces distinctive patterns that reveal structural information. Monocrystalline materials generate sharp bright spots, polycrystalline specimens create concentric rings, and amorphous materials produce diffuse rings or broad halos. Elastically scattered electrons, which retain their original energy, are essential for high-resolution imaging and are responsible for the electron diffraction patterns used to extract crystallographic information in both TEM and STEM techniques. Inelastic scattering involves significant energy transfer from incident electrons to the specimen, primarily through electron-electron interactions.161 This energy transfer can excite or ionise electrons within the specimen, generating various secondary signals. When an inner- shell vacancy is filled by an outer-shell electron, characteristic X-rays may be emitted. Low- energy electrons ejected from outer shells form secondary electrons, while Auger electrons are emitted when excess energy from electronic transitions transfers to another electron instead 36 3.2 Transmission Electron Microscopy of being released as X-rays. Most of these signals can be detected in both SEM and TEM/STEM systems, with characteristic X-rays commonly analysed using Energy Dispersive X-ray Spectroscopy (EDS) and secondary electrons utilised for topographical imaging in both techniques. Inelastically scattered electrons themselves provide valuable analytical information through electron energy loss spectroscopy (EELS), revealing chemical bonding states, oxidation states, elemental composition, and local electronic structure.124 EELS is predominantly available in TEM/STEM configurations and not implemented in conventional SEM systems due to the need for transmitted electrons and specialised energy filtering mechanisms. Unlike elastic scattering, inelastic interactions are typically incoherent, as energy exchange disrupts the phase relationship of the electron wave, reducing its ability to contribute to coherent imaging and diffraction contrast. 3.2 Transmission Electron Microscopy A conventional TEM column consists of several key components arranged from top to bottom: an electron gun, condenser lens system, objective lens and imaging optics, and a detection system159. While the scattering effects described earlier determine how electrons interact with the specimen to produce contrast and generate analytical signals, the role of the instrument components is to precisely form, shape, focus, and deliver the electron beam onto the specimen and subsequently to collect and analyse the resulting signals with high spatial and energy resolution. Each component plays a critical role in maintaining beam coherence, optimising resolution, and enabling the range of imaging and spectroscopic capabilities available in modern TEM systems. 3.2.1 Electron Gun The electron gun, positioned at the top of the TEM column, generates the primary electron beam that is subsequently shaped and directed onto the specimen. There are two main types of electron sources used in TEM:13 • Thermionic sources (W of LaB6): Electrons are emitted by thermal excitation when heated to high temperature • Field Emission Guns (FEGs): Electrons are extracted from a sharp tungsten tip via quantum tunnelling, enabled by a strong electric field applied between the tip and extraction anode 3.2 Transmission Electron Microscopy 37 An important parameter for comparing electron sources is the brightness (𝛽), defined as the current density emitted per unit solid angle of the electron source. This can be expressed as:159 𝛽 = 4𝐼𝑒 𝜋2𝐷0 2𝛼0 2 = 4𝐽𝑒 𝜋2𝛼0 2 Equation 1 where 𝐼𝑒 is the total emitted current, 𝐷𝑜 is the source size, 𝛼0 is the beam divergence semi- angle, and 𝐽𝑒 is the current density. A higher brightness enables the generation of a finer, more intense probe, improving both imaging and spectroscopic capabilities. For this research, a ThermoFisher Spectra-300 microscope equipped with a Schottky FEG was used, operating at an acceleration voltage of 300 kV. This choice provided an optimal balance of high brightness, excellent beam stability, and low noise. These attributes are essential for high-resolution imaging and analytical techniques such as STEM-EELS, where high spatial coherence and beam current are critical for investigating the electronic structure of NMC811 materials. Table 3.1: Characteristics of the principal electron sources at an acceleration voltage of 100 kV.13 W Thermionic LaB6 Thermionic Schottky FEG Cold FEG Brightness / A/m2sr 1010 5 x 1011 5 x 1012 1013 Current density A/m2 5 102 105 106 Crossover size / n > 105 104 15 3 Lifetime / h 100 1000 >5000 >5000 Table 3.1 summarises the key performance metrics of typical electron sources at an acceleration voltage of 100 kV. Although both thermionic and FEG sources may use tungsten, the emission mechanism and resulting beam properties differ significantly. The W thermionic source produces relatively low brightness due to its larger crossover size and lower current density. In contrast, the FEG, particularly the cold FEG, achieves brightness values up to 1013 A m-2 sr, owing to its small source size and efficient electron extraction via tunnelling. 38 3.2 Transmission Electron Microscopy 3.2.2 Condenser Lens Below the electron gun is the condenser lens system, which controls the illumination conditions at the specimen. It typically consists of two or more electromagnetic lenses and associated apertures, and serves to:159 • Demagnify the electron source to form a smaller, brighter probe • Control the beam diameter and beam current at the specimen plane • Regulate the convergence angle (𝛼) • Setting illumination modes: Enables switching between parallel and convergent beam configurations Figure 3.2: The two lens condenser system. The spot of size s1 at the gun crossover (G) is demagnified to s2 by the first condenser lens C1. The second condenser lens C2 is used to focus the beam. (A), (B) and (C) show underfocused, focused and overfocused beams respectively. The convergence angle 𝛼 is controlled by the condenser aperture. In a standard two-lens system, the first condenser lens (C1) demagnifies the electron source at the gun crossover (G), reducing the spot size (s2). Increasing C1 excitation yields a smaller, more coherent probe with lower current, while decreasing it results in a larger, higher-current A B C 3.2 Transmission Electron Microscopy 39 probe. The second condenser lens (C2) focuses or defocuses the beam, adjusting convergence at the specimen. As shown in Figure 3.2, this enables three illumination modes: A. Underfocused - broad illumination for low magnification image, B. Focused - tight beam for analytical applications C. Overfocused - convergent probe for techniques like convergent beam electron diffraction. The condenser aperture, positioned below C2, further shapes the beam by limiting the convergence semi-angle (α), reducing spherical aberrations, regulating beam current, and enhancing spatial coherence by filtering out high-angle electrons. The third condenser lens (C3), available in the Spectra 300 system used in this study, provides independent control of beam convergence and current, allowing fine tuning of probe conditions without altering spot size. This is essential for analytical modes like STEM-EELS and for maintaining stable illumination during switching between TEM and STEM. C3 is also critical for aberration- corrected imaging, where precise control over probe formation is required to achieve sub- ångström resolution. 3.2.3 Objective Lens and Imaging System After passing through the specimen, the transmitted electrons are collected by the objective lens, which forms either a real-space image in the image plane or a diffraction pattern (DP) in the back focal plane. These outputs are further magnified by a series of intermediate and projector lenses, allowing the final image or diffraction pattern to be visualised on a screen or detector. This process is illustrated in Figure 3.3. Like all electromagnetic lenses, the objective and projector systems suffer from spherical and chromatic aberrations, as well as astigmatism, which degrade spatial resolution162. These can be mitigated using stigmators and, in modern instruments, aberration correctors. Additional components such as beam deflectors, scanning coils, and drift correctors further support precise beam positioning and image stability, which are essential for high-resolution imaging and analytical applications. 40 3.2 Transmission Electron Microscopy Figure 3.3: Schematic of objective and intermediate lens operation in TEM. (a): When the intermediate lens focuses on the image plane of the objective lens, a magnified real-space image is formed. (b): When refocused on the back focal plane, a diffraction pattern is projected. The position of apertures and the lens configuration determine whether the output is an image or a diffraction pattern. 3.2.4 Detectors There are several ways to collect electron signals in both TEM and STEM modes. The most commonly used detector is the charge-coupled device (CCD), which records images, spectra, and diffraction patterns. 159 CCDs consist of metal-insulator-silicon arrays that store charge in each pixel, proportional to incoming signal intensity. As CCDs are inherently sensitive to photons rather than electrons, a scintillator screen is placed before the sensor to convert incident electrons into visible light, which is then detected and digitised by the CCD. CCDs are also widely used for two-dimensional parallel collection in EELS and energy- filtered imaging. Other detectors, such as bright-field (on-axis) and annular dark-field (off- axis) detectors, are used in STEM mode to collect scattered electrons at different angles. These will be discussed in more detail in the following section, as they are integral to STEM image formation. (a) (b) 3.3 Scanning Transmission Electron Microscopy 41 3.3 Scanning Transmission Electron Microscopy 3.3.1 Historical Development and Principles Scanning Transmission Electron Microscopy (STEM) integrates the strengths of Scanning Electron Microscopy (SEM) and Transmission Electron Microscopy (TEM), offering high signal intensity and sub-nanometre spatial resolution. Unlike SEM, which primarily detects secondary or backscattered electrons for surface imaging, STEM captures transmitted and scattered electrons, enabling atomic-resolution imaging and advanced analytical techniques.163,164 The concept of STEM was introduced by Manfred von Ardenne in 1937,165 who proposed placing the imaging lens before the specimen to reduce aberrations. However, due to the lack of a suitable field emission source, early systems produced noisy images with resolutions around 10 nm. Nearly 30 years later, Albert Crewe166 demonstrated the potential of STEM using a cold FEG, achieving resolutions down to 5 Å and paving the way for modern STEM instruments. In STEM mode, a finely focused electron probe is raster-scanned across the specimen whilst detecting various signals at each probe position, including transmitted electrons, inelastically scattered electrons, secondary electrons, and X-rays.164 The spatial resolution in STEM is fundamentally limited by the probe diameter, necessitating a small, bright, and coherent beam, which is typically achieved using a field emission gun. 3.3.2 Probe Formation and Beam Scanning In STEM mode, the electron beam is converged into a finely focused probe that is systematically scanned across the specimen in a raster pattern. At each scan point, signals are collected and mapped to form high-resolution images or spectrum images with nanometre to atomic precision.164 The probe itself is a demagnified image of the electron source, and achieving a small, bright probe is essential for high spatial resolution. Several key factors influence probe formation: • Source Brightness: A higher brightness electron source increases beam current within a small probe size, improving signal intensity and enabling better contrast in high- resolution imaging. • Lens Aberrations: Among geometric aberrations, spherical aberration (Cs) has historically been the dominant limiting factor for probe size. Rays travelling at large 42 3.3 Scanning Transmission Electron Microscopy angles to the optic axis are focused at different positions from those near the axis, causing blurring at high resolution. • Chromatic Aberration: Caused by electron energy spread, chromatic aberration causes electrons of different energies to focus at varying depths. This is reduced using field emission guns (FEGs), which produce narrow energy distributions. The introduction of aberration-corrected STEM instruments has fundamentally transformed electron microscopy capabilities.167 These sophisticated corrector systems effectively compensate for spherical aberrations through multipole lenses arranged in specific configurations, enabling the formation of sub-ångström electron probes with unprecedented spatial resolution. This technological advancement facilitates atomic-resolution imaging and spectroscopy that was previously unattainable with conventional instruments. Beyond enhanced spatial resolution, aberration correction substantially increases the beam current density within the probe, leading to significant improvements in signal-to-noise ratio and image contrast. These benefits are particularly crucial for applications requiring stringent dose control or advanced analytical techniques such as EELS or EDS, where signal intensity directly impacts spectral quality and chemical sensitivity.168 Once formed, the probe is deflected across the sample surface using scan coils placed above the objective lens. An objective aperture is typically inserted to restrict the maximum illumination angle (𝛼), helping to control beam convergence and reduce spherical aberrations. This scanning mechanism allows precise localisation of signals to individual pixels, enabling both structural and chemical mapping at the nanoscale.167 3.3.3 Detection system and signal collections A significant advantage of STEM is its ability to collect signals over a wide angular range using multiple detectors positioned beneath the specimen. The STEM detection geometry is illustrated in Figure 3.4, showing how electrons scattered at different angles are captured by distinct detectors to provide complementary imaging modes:169 • A bright-field (BF) detector captures low-angle, forward-scattered electrons close to the optic axis (𝜃3 < ~10 mrad). • An annular dark-field (ADF) detector surrounds the transmitted beam and collects elastically scattered electrons at intermediate angles (𝜃2 ~10 - 50 mrad). • At even higher scattering angles ( 𝜃1 > ~50 mrad), high-angle annular dark-field (HAADF), often referred to as Z-contrast imaging, is formed. A bright-field (BF) detector captures low-angle, forward-scattered electrons close to the optic axis. 3.3 Scanning Transmission Electron Microscopy 43 HAADF imaging is particularly valuable because the signal intensity is approximately proportional to the atomic number, I∝Z1.8 making heavier elements appear brighter.169 BF images display vacuum regions as bright (due to direct transmission of the beam), whereas ADF and HAADF images show vacuum as dark, since the transmitted beam is excluded. These contrast mechanisms are inherently complementary and provide rich structural and compositional information from a single scan. Figure 3.4: (a) Schematic of STEM detection geometry showing typical detector positions for EDX, EELS, and HAADF signal collection and (b) Angular ranges for scattered electrons captured by BF, ADF, and HAADF detectors, enabling complementary contrast based on scattering angle and atomic number.170 Another key advantage of STEM over conventional TEM is the availability of space beneath the specimen, since post-specimen imaging lenses are not required. This enables the integration of advanced analytical systems such as energy-dispersive X-ray spectroscopy (EDS) and electron energy-loss spectroscopy (EELS).163 Most modern STEM instruments incorporate a parallel EELS system, consisting of a magnetic spectrometer and a CCD detector, allowing for high-resolution spectral mapping in tandem with imaging. The ability to operate multiple detectors simultaneously makes STEM a powerful platform for correlative analysis, combining atomic-resolution imaging with precise chemical and electronic structure characterisation in a single experiment. (a) 44 3.3 Scanning Transmission Electron Microscopy 3.3.4 Sampling in STEM The spatial resolution in STEM imaging and spectroscopy depends on both the effective probe size (𝑑eff) and the pixel size used during scanning. Three sampling regimes exist, as illustrated in Figure 3.5. Figure 3.5: Sampling regimes in STEM: (a) exact sampling, (b) undersampling, and (c) oversampling, based on the relationship between probe size (𝑑eff) and pixel size For optimal results, the choice between exact sampling and oversampling depends on the specific application. Exact sampling is theoretically ideal for general imaging as it maximises efficiency by sampling each region exactly once. For spectroscopic applications like EELS or EDS, oversampling is preferred as it would:163 • Improve signal-to-noise ratio through increased sampling density • Enables post-acquisition processing techniques like noise reduction and rebinning • Provides redundant data that helps compensate for drift or instabilities • Ensures no fine features are missed between sampling points However, oversampling increases both acquisition time and electron dose, which can be problematic for beam-sensitive materials like NMC811. The additional electron exposure must be carefully balanced against the analytical benefits. Exact Sampling Pixel size = 𝑑eff, providing complete coverage without overlap Undersampling Pixel size > 𝑑eff, leaving gaps between sampled points and potentially missing fine details. Oversampling Pixel size < 𝑑eff, causing beam overlap between adjacent pixels. 3.4 Electron Energy Loss Spectroscopy 45 3.4 Electron Energy Loss Spectroscopy Electron Energy-Loss Spectroscopy (EELS) was selected as the primary analytical technique for this research due to its exceptional sensitivity to electronic structure changes in transition metal oxides such as NMC811. In EELS, high-energy electrons undergo inelastic scattering, losing discrete amounts of energy as they excite atomic electrons from occupied to unoccupied states.124 This technique provides crucial insights into the oxidation states and chemical bonding environments of the constituent elements (Ni, Mn, Co, and O) within the cathode material, enabling direct correlation between electrochemical cycling and local electronic structure evolution. Figure 3.6: A schematic representation of the allowed electronic transitions from occupied atomic orbitals (core shells) to unoccupied states, as a function of increasing energy. The K, L, M, N, and O shells correspond to principal quantum numbers (n = 1, 2, 3, 4, 5), with sublevels split by orbital angular momentum (l) and spin-orbit coupling (j). Transitions labelled K, L₂,₃, M₂,₃, etc., represent characteristic core-loss edges observable in EELS, and are used for elemental identification and oxidation state analysis. The direction of transitions follows dipole selection rules (Δl = ±1).159 As illustrated in Figure 3.6, energy losses in EELS correspond to transitions from core electronic levels to unoccupied states above the Fermi level. The resulting spectrum is typically divided into two regions: the low-loss region (<50 eV), which includes the zero-loss peak (ZLP) and plasmon excitations, and the core-loss region (>50 eV), where sharp ionisation edges appear due to core-level excitations. 46 3.4 Electron Energy Loss Spectroscopy For NMC811 cathode materials, several spectral features are particularly important:159 • Transition Metal L₂,₃ Edges: The L₂,₃ edges of Ni, Mn, and Co correspond to 2p → 3d transitions and exhibit distinct white-line features. These white-line intensities, energy shifts, and fine structures are collectively referred to as energy-loss near-edge structure (ELNES). The ELNES are sensitive to the oxidation state and coordination environment of the absorbing atom. An increase in oxidation state typically leads to a shift of the L₃ edge to higher energy and a change in the L₃,₂ intensity ratio due to increased 3d orbital vacancy. • Oxygen K-Edge: The O K-edge arises from 1s → 2p transitions, where the unoccupied O 2p states are hybridised with the 3d states of neighbouring transition metals. The pre-edge and near-edge features of the O K-edge reflect both the oxygen electronic structure and the covalency of the metal-oxygen bonds. These features are especially useful for tracking redox processes in NMC811, where oxygen may participate in charge compensation during delithiation. • Low-Loss Features: The low-loss region (<50 eV), which includes the zero-loss peak (ZLP) and plasmon excitations, provides valuable information about electronic density and can be used to estimate local specimen thickness, critical for quantitative analysis. 3.4.1 Multiple Scattering Considerations The reliability of EELS analysis critically depends on the specimen thickness, which directly influences the probability distribution of scattering events. As electrons traverse a specimen, they may interact with atoms multiple times before exiting, complicating the direct interpretation of imaging and spectroscopic data. The probability and extent of multiple scattering events are directly related to specimen thickness, which must be carefully controlled and characterised for accurate analysis.124 Any scattering event in particle physics can be described by a cross-section (𝜎) and a mean free path (𝜆)124. In electron microscopy, the cross-section represents the likelihood of a certain type of scattering occurring, influenced by factors such as sample thickness, composition, beam energy and incidence angle. The mean free path (𝜆) describes the average distance that an electron travels between successive scattering events, and is inversely proportional to σ. For high-energy electrons in TEM, 𝜆 is typically in the order of tens to hundreds of nanometres, depending on specimen composition and microscope parameters. 3.4 Electron Energy Loss Spectroscopy 47 An estimation of the inelastic mean free path can be calculated using the dipole formula derived by Egerton:124 𝜆 = 106𝐹 ( 𝐸0 𝐸𝑚 ) ln ( 2𝛽𝐸0 𝐸𝑚 ) Equation 2 Where F the relativistic factor, 𝐸0 the beam energy in keV, 𝛽 is the collection semi-angle in radians, and 𝐸𝑚 is the mean energy loss per inelastic scattering event. The parameters used in this work for NMC811 are summarised in Table 3.2. A beam energy of 300 kV and a collection semi-angle of 17 mrad were chosen to optimise spatial resolution and minimise beam broadening while maintaining a sufficiently long inelastic mean free path (𝜆) to limit plural scattering. These parameters also ensure reliable detection of transition metal L2,3 edges and ELNES features in NMC811 Table 3.2: Microscope parameters and material constants used for multiple scattering analysis, based on a beam energy (𝐸0) of 300 kV, a collection semi-angle of 17 mrad, and a relativistic factor (F) of 0.514. 𝑍𝑒𝑓𝑓 is the effective atomic number, 𝐸𝑚 is the characteristic energy loss per inelastic scattering event, and 𝜆 is the inelastic mean free path. For NMC811, 𝜆 values are compared across three structural phases: layered Li(Ni0.8Mn0.1Co0.1)O2, spinel Li(Ni0.5Mn1.5)O4, and rock-salt Ni2O. Additional reference materials, including hybrid perovskite, Cs0.1[CH(NH2)2]0.9Pb(I0.955Br0.045)3, along with WO3, SiO2, and SnO2 are included for comparison to NMC811. 𝒁𝒆𝒇𝒇 𝑬𝒎 / eV 𝝀 / nm NMC811 (Layered) 13.9 19.6 133.2 NMC811 (Spinel) 14.2 19.7 119.6 NMC811 (Rock-salt) 19.9 22.3 132.5 Hybrid Perovskite 37.8 28.1 98.6 WO3 34.0 27.0 101.8 SiO2 10.2 17.6 146.1 SnO2 27.5 25.1 108.5 48 3.4 Electron Energy Loss Spectroscopy To estimate the mean loss per inelastic scattering, the effective atomic mass number (𝑍𝑒𝑓𝑓) must be calculated. For complex specimens comprising multiple elements, the 𝑍𝑒𝑓𝑓 can be estimated as follows:124 𝑍𝑒𝑓𝑓 ≈ ∑ 𝑓𝑖𝑍𝑖 1.3 𝑖 ∑ 𝑓𝑖𝑍𝑖 0.3 𝑖 Equation 3 𝐸𝑚 ≈ 1.6 𝑍𝑒𝑓𝑓 0.36 (𝑒𝑉) Equation 4 With 𝑓𝑖 the atomic molar fraction and 𝑍𝑖 is its atomic number Since TEM lamellae typically have thicknesses comparable to or greater than the calculated 𝜆, multiple scattering becomes significant. As each scattering event occurs independently, the Poisson distribution provides an effective model for calculating the probabilities of n scattering events:124 𝑃𝑛 = 1 𝑛! 𝑡𝑛 𝜆 𝑒𝑥𝑝−𝑡/𝜆 Equation 5 Where n is the number of scattering events (0, 1, or more for multiple scattering), 𝑡 is the sample thickness and 𝜆 is the mean free path calculated. Table 3.3 quantifies inelastic scattering events in electron energy-loss spectroscopy (EELS) as a function of specimen thickness, which critically influences both data quality and spectral interpretation for 𝑡/𝜆 = 1, 0.75, 0.5 and 0.3. Each scattering regime has distinct implications for EELS analysis of NMC811. P0 represents electrons that traverse the specimen without energy loss, contributing to the zero-loss peak and serving as a reference for energy calibration and resolution assessment. These electrons are essential for accurate spectral alignment despite containing no core-loss information. P1 denotes single scattering events, which are ideal for EELS as they carry interpretable information about elemental composition and electronic structure, including the oxidation states of transition metals such as Ni, Mn and Co, as well as oxygen. As thickness increases, the probability of P1 eventually declines, giving rise to plural scattering events (P2 and P3) that introduce convolution artefacts and overlapping spectral features, degrading the clarity of fine structure and complicating oxidation state determination124. 3.4 Electron Energy Loss Spectroscopy 49 Table 3.3: The probabilities for zero (P0), single (P1), double (P2), and triple (P3) scattering events at selected sample thicknesses and their corresponding 𝑡/𝜆 values, at 300 keV and 17 mrad Scattering Events 𝒕 = 𝟏𝟑𝟓 𝒏𝒎 𝒕/𝝀 = 𝟏 𝒕 = 𝟏𝟎𝟎 𝒏𝒎 𝒕/𝝀 = 𝟎. 𝟕𝟓 𝒕 = 𝟔𝟓 𝒏𝒎 𝒕/𝝀 = 𝟎. 𝟓 𝒕 = 𝟒𝟎 𝒏𝒎 𝒕/𝝀 = 𝟎. 𝟑 No scattering, 𝑃0 36 % 47 % 61 % 74 % Single scattering, 𝑃1 37 % 35 % 30 % 22 % Double scattering, 𝑃2 19 % 13 % 7 % 3 % Triple scattering, 𝑃3 6 % 3 % 1 % 0 % The signal-to-noise ratio (SNR) in EELS varies nonlinearly with thickness. Very thin specimens (𝑡/𝜆 < 0.3) yield clean spectra with minimal background from multiple scattering but suffer from poor signal intensity due to the low occurrence of inelastic events. A more favourable regime is found at moderate thickness (t/λ ≈ 0.5), where P1 maintains adequate probability (~30%) while multiple scattering remains limited. In thicker specimens (t/λ > 0.5), although the absolute P1 value may still be significant, spectral quality begins to degrade due to increased contributions from P2 and P3, often necessitating deconvolution techniques. These procedures, while potentially helpful, risk amplifying noise, introducing artefacts, and distorting subtle fine structure features, and are particularly problematic when interpreting delicate ELNES features associated with oxidation state changes in NMC811. Through systematic analysis of scattering probabilities presented in Figure 3.7, an optimal thickness range of t/λ ≈ 0.5 was established for this research. At this thickness, single scattering events maintain sufficient probability (P₁ ≈ 30%) to generate adequate signal intensity whilst multiple scattering contributions remain manageable at 8%, which is more than a twofold reduction from t/λ ≈ 0.75, without significant loss in P₁. This represents the best compromise between signal strength and spectral clarity, ensuring reliable interpretation of ELNES features whilst minimising deconvolution artefacts. This thickness optimisation was particularly important for the beam-sensitive NMC811 materials investigated in this work. Conventionally, TEM lamellae are prepared using focused ion beam-scanning electron microscopy (FIB-SEM) to thicknesses of 100-150 nm. Achieving the optimised thickness of approximately 65 nm required further refinement of FIB-SEM preparation techniques, as detailed in Section 3.5.3. This careful thickness control enabled acquisition of high-quality EELS data whilst minimising both plural scattering artefacts and electron beam-induced damage, essential for accurate oxidation state determination in these 50 3.4 Electron Energy Loss Spectroscopy electrochemically active materials, without necessitating deconvolution procedures that could potentially compromise the integrity of fine spectral features critical for characterising the electronic structure of degraded NMC811 regions. Figure 3.7: Probability of multiple inelastic scattering events against sample thickness and their corresponding 𝑡/𝜆 in vertical dashed lines. Each curve represents the probability of electrons undergoing zero (P0), single (P1), double (P2), or triple (P3) scattering events of NMC811 3.4.2 EELS Acquisition in this Thesis EELS data acquisition in this research utilises scanning transmission electron microscopy (STEM) to achieve spatially resolved spectroscopic imaging with nanometre precision. As illustrated in Figure 3.8, the focused electron probe systematically traverses the specimen in a raster pattern, generating a comprehensive three-dimensional dataset known as a spectrum image, comprising two spatial dimensions (x, y) and one spectral dimension (ΔE)124. This approach enables direct correlation between local structural features and electronic properties, critical for characterising heterogeneous degradation processes in cycled NMC811 cathodes. The electron detection pathway in EELS represents a sophisticated sequence of electron- optical interactions. Following inelastic scattering events within the specimen, transmitted electrons enter a post-column energy-loss spectrometer where a magnetic prism applies a perpendicular magnetic field, dispersing electrons according to their kinetic energies. This magnetic sector effectively separates electrons that have experienced different energy losses during specimen interaction. The dispersed electron beam subsequently passes through a 3.4 Electron Energy Loss Spectroscopy 51 series of multipole lenses that correct aberrations and optimise energy resolution before projection onto a high-sensitivity detector159. In the Gatan Quantum GIF system employed for this research, a charge-coupled device (CCD) camera with high detective quantum efficiency converts the spatially separated electrons into digital signals, enabling simultaneous acquisition across the energy spectrum with exceptional signal-to-noise characteristics. EELS experiments were conducted using a Gatan Quantum GIF spectrometer attached to a ThermoFisher Spectra-300 transmission electron microscope operating at 300 kV. The spectra were acquired in DualEELS mode, facilitating concurrent collection of low-loss and core-loss regions, an essential capability for accurate thickness determination and plural scattering correction. An energy resolution of approximately 1.5 eV was achieved with a dispersion of 0.3 eV/channel, providing sufficient spectral detail to resolve fine structural features in the transition metal L₂,₃ edges and oxygen K-edge. Figure 3.8: A converged electron probe is rastered across the specimen in STEM mode. At each probe position, the transmitted electrons are collected using a post-column EELS spectrometer, which records an energy-loss spectrum. Simultaneously, high-angle scattered electrons can be detected using a dark-field (DF) or HAADF detector. The resulting data forms a three- dimensional spectrum image (x, y, ΔE), where each pixel contains a full EELS spectrum. This data cube enables spatially resolved chemical and electronic structure analysis. 171 52 3.4 Electron Energy Loss Spectroscopy Each EELS spectrum was acquired with a dwell time of 0.5 seconds per pixel, using a pixel size of 1 nm, providing spatially resolved chemical information. The beam current was maintained at approximately 130 pA, and spectra were collected with a collection semi-angle of 17 mrad. These parameters were selected to provide sufficient signal-to-noise for quantitative analysis of the transition-metal 𝐿2,3 edges and the O K-edge, while minimising beam-induced artefacts. To validate the dose calculations and confirm that the chosen acquisition window did not induce structural degradation, systematic beam-damage tests were performed on pristine NMC811. High-resolution TEM images were acquired from the same region before and immediately after controlled electron-beam exposure at different dose levels (Figure 3.9). Under a high-dose condition of 500 pA for 1 s (electron dose ≈ 3.12 × 109 e- over the illuminated area), the post-exposure image shows clear evidence of beam damage, including the formation of a local amorphous region visible as a dark spot and a marked reduction in lattice periodicity. This is accompanied by a corresponding loss of crystallinity in the Fourier transform, with weakened and less well-defined spots relative to the pre-exposure image. In contrast, under a lower-dose condition of 150 pA for 0.5 s (electron dose ≈ 4.68 × 108 e-), no detectable change was observed within the sensitivity of this test. The lattice fringes remain clearly resolved and the Fourier transform retains sharp diffraction spots, indicating preservation of the layered structure. These results confirm that the acquisition conditions used in this thesis provide adequate spectral quality for quantitative EELS analysis while remaining below the onset of observable beam damage in NMC811. Finally, the selected microscope settings are consistent with those commonly used in prior STEM-EELS investigations of Ni-rich layered oxide cathodes, where comparable accelerating voltages, probe currents, and exposure times are employed to resolve core-loss edges and ELNES features while limiting beam-induced reduction and structural disorder. 26,172 3.4 Electron Energy Loss Spectroscopy 53 Figure 3.9: Beam-damage assessment for EELS acquisition conditions in NMC811. High- resolution TEM images and corresponding FFTs acquired before and after electron-beam exposure at two dose levels. Top: High-dose condition (500 pA, 1 s; ~3.12×109 e-) shows beam- induced damage with local amorphisation and altered FFT patterns. Bottom: Lower-dose condition (150 pA, 0.5 s; ~4.68×108 e-) shows no detectable structural change with preserved lattice fringes and unchanged FFT spots. 54 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB- SEM) Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM) is a powerful dual-beam platform that combines a focused ion beam for material removal with a scanning electron microscope for high-resolution imaging. Initially developed for the microelectronics industry nearly four decades ago, FIB systems were used for tasks such as circuit editing and defect analysis. Over time, their applications have expanded significantly to include site-specific sample preparation for transmission electron microscopy (TEM), microstructural analysis, and prototype nanofabrication173. FIB-SEM was employed as a critical sample preparation technique in this research, enabling the production of high-quality specimens for subsequent Electron Energy Loss Spectroscopy (EELS) analysis as well as providing complementary 3D microstructural information. While EELS served as the primary analytical technique for investigating electronic structure changes in cycled NMC811 cathodes, FIB-SEM was essential for preparing electron-transparent specimens of controlled thickness, a critical parameter for obtaining reliable EELS data with minimal plural scattering artefacts. This section details the FIB-SEM instrumentation, sample preparation techniques, and analytical protocols developed and utilised throughout this investigation. 3.5.1 Instrumentation and Operating Principles The experimental work was conducted using two complementary FIB-SEM systems: 1. An FEI Helios NanoLab dual-beam system for TEM lamella preparation 2. A Zeiss CrossBeam 540 dual-beam Ga+ FIB-SEM for cross-sectional imaging and 3D segmentation 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) 55 Figure 3.10: Schematic of modern dual-beam columns FIB systems. The electron beam and an ion beam allow for visualisation of the sample and deposition/milling in the region of interest. Figure from Carl Zeiss.174 Early FIB systems used a single ion column equipped with a gallium ion source for both milling and imaging.175 However, these systems presented several limitations: ion imaging often introduced damage, the beam lacked the ability to generate secondary signals such as X-rays or backscattered electrons, and there was limited visual feedback during milling. To address these issues, modern FIB-SEM instruments operate on a dual-beam design principle, as shown in Figure 3.10, featuring an SEM column for non-destructive electron imaging and an ion column positioned at a typical angle of 52 ° to the electron column. This configuration allows simultaneous imaging and milling of the same region with exceptional precision. The SEM column employs electromagnetic lenses to focus an electron beam for imaging, while the FIB column uses electrostatic lenses to focus a heavy ion beam, typically gallium, for material removal.175 Beam conditions were precisely controlled through a system of apertures that regulate current and spot size. For high-resolution imaging, low currents in the picoampere (pA) range were utilised, while higher currents (up to tens of nanoamperes) were employed for rapid bulk milling operations. The acceleration voltage of the ion beam was adjusted according to specific requirements, with higher voltages for efficient material removal and lower voltages for fine polishing to minimise beam-induced damage. 56 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) 3.5.2 TEM Lamella Preparation For transmission electron microscopy (TEM) and Electron Energy Loss Spectroscopy (EELS) analysis, high-quality electron-transparent lamellae were prepared using the in-situ lift-out technique.173 This method was selected for its ability to target specific regions of interest with nanometer precision, a critical requirement for analysing the heterogeneous microstructure and electronic properties of NMC811 cathode materials. The preparation of exceptionally thin specimens was particularly important for EELS analysis, where sample thickness directly impacts the quality of spectroscopic data by minimising multiple scattering effects. Figure 3.11: Sequential SEM (orange border) and FIB (purple border) images illustrating the in-situ lift-out process for TEM lamella preparation. 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) 57 Figure 3.11 illustrates the preparation workflow, which involves: 1. Protective Layer Deposition: A dual-layer protective coating was applied, beginning with electron beam-assisted platinum deposition (~500 nm) followed by ion beam- assisted deposition (1-3 µ𝑚). This protective measure prevented gallium implantation and reduced surface damage during subsequent milling operations. 2. Trenching: Trenches were milled on both sides of the protected region to a depth of several microns, isolating a central strip of material containing the region of interest. 3. Undercutting: Additional cuts were made to form a "J-shape," leaving the lamella connected to the bulk sample at a single point. 4. Lift-Out: A micromanipulator was attached to the free end of the lamella using ion- assisted platinum deposition, after which the final attachment point was milled away to completely free the sample. 5. Mounting: The extracted lamella was transferred to a copper half-grid and secured using platinum deposition. 6. Final Thinning: The lamella was progressively thinned to ~100 nm using decreasing ion beam currents and appropriate tilting strategies to minimize curtaining artifacts and beam damage. 3.5.3 Modified Mounting Approach for Ultrathin Lamellae for EELS A significant methodological advancement in this work was the development of a modified mounting approach to address the challenges presented by the brittle nature of NMC811 materials. Conventional side-mounting techniques limited thinning to approximately 100 - 150 nm due to mechanical instability during final polishing, which was insufficient for high- resolution EELS analysis. For accurate EELS characterisation, especially when analysing fine chemical variations at interfaces or determining oxidation states in transition metals, sample thicknesses below 70 nm are often necessary to minimise multiple scattering effects and optimise energy resolution in NMC811. 58 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) Figure 3.12: Optical image of (a) a Cu half-grid from Agar Scientific showing multiple mounting positions (labelled A-C), and (b) close-up view highlighting finger B for conventional side-mounted lamella and finger C for the modified M-slot mounting. Red boxes indicate the regions where lamellae are typically attached using Pt deposition176 To overcome this limitation, a custom mounting configuration was implemented using copper half-grids with an M-shaped slot design. Rather than attaching the lamella along the grid edge, it was positioned across the top of the central M-slot and secured at both bottom corners with platinum deposition, as shown in in Figure 3.12. This approach provided enhanced mechanical stability by anchoring the lamella directly between grid posts, substantially reducing flexing during thinning operations. The modified mounting technique enabled reliable preparation of ultrathin lamellae down to ~40 nm, improved control during low-kV final polishing steps, enhanced sample stability for high-resolution analytical techniques, and reduction in mechanical failure rates during preparation. The modified mounting and thinning procedure is shown in Figure 3.13. Figure 3.13: SEM (orange border) and FIB (purple border) images showing the modified lamella mounting and thinning strategy for improved mechanical stability and ultrathin TEM sample preparation (a) (b) 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) 59 3.5.4 FIB-SEM for 3D-segmentation For comprehensive microstructural characterization of cycled cathodes, FIB-SEM tomography was performed using the Zeiss CrossBeam 540 system. The workflow, as illustrated in Figure 3.14, for these experiments included: 1. Sample Preparation: Cycled cathodes were washed using dimethyl carbonate and mounted onto aluminum stubs using carbon tape. Samples were stored in an argon glove box prior to analysis to prevent atmospheric contamination. 2. Protective Layer Application: A 1 μm thick platinum layer was deposited over an area of 20 μm × 30 μm to protect the surface region during subsequent milling operations. 3. Volume Preparation: Initial rough milling of trenches was conducted using a 30 nA gallium ion beam at 30 kV to prepare a volume of 20 μm × 30 μm × 30 μm. A reference pattern was milled into the sample to facilitate drift compensation during the sequential milling and imaging process. 4. Drift Correction Marker: A fiducial marker was milled adjacent to the region of interest to enable alignment and compensate for lateral drift during sequential milling and imaging steps. 5. Surface Cleaning and Slice Removal: Redeposited material from the rough milling stage was removed using a 1.5 nA gallium ion beam. Cross-sectional slices of approximately 60 nm in nominal thickness were sequentially removed using the same beam conditions. 6. High-Resolution Imaging: Images were acquired in electron immersion mode with a pixel size of 10 × 10 nm² using a 2 kV accelerating voltage and 300 pA beam current to optimize surface detail and compositional contrast. 60 3.5 Focused Ion Beam – Scanning Electron Microscopy (FIB-SEM) Figure 3.14: Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) workflow for 3D tomography of cycled NMC811 cathodes. (a) Schematic representation of the FIB-SEM slice- and-view process, showing the geometry of slice removal relative to the ion and electron beam axes. The protective platinum layer and fiducial marker are also indicated. Figure adapted from Liu et al.177 (b) SEM image of the sample following initial trench milling using a 30 nA gallium ion beam, showing redeposition around the milled volume. (c) Further polishing with a 1.5 nA ion beam reveals the detailed microstructure; vertical streaks (curtaining artefacts) are visible due to milling inhomogeneity. (d) Final high-resolution cross-sectional image used for segmentation, illustrating the distinct phases of the electrode microstructure. Scale bars in (b-d) represent 5 𝜇𝑚. 3.6 Artefact Mitigation Strategies 61 3.6 Artefact Mitigation Strategies Several potential artefacts are associated with FIB-based sample preparation, including surface amorphisation, redeposition of sputtered material, gallium implantation, and contamination from protective coatings175. To minimise these effects, multiple mitigation strategies were systematically implemented throughout the sample preparation workflow. As illustrated in Figure 3.15, three primary artefacts commonly encountered in FIB-SEM imaging of NMC811 include curtaining, shine-through, and charging. Curtaining artefacts, caused by uneven sputtering rates during ion beam milling, appear as vertical streaks or bands that obscure important microstructural features. To reduce this issue during TEM lamella preparation, electron beam-assisted platinum deposition was employed as the initial protective step, forming a sacrificial layer to shield the surface from direct ion beam exposure. Final thinning and polishing were performed using progressively lower ion beam energies, concluding with low-energy cleaning steps (≤5 kV) to minimise surface damage depth. For ultrathin lamellae intended for high-resolution EELS, a modified mounting strategy was adopted to improve mechanical stability and prevent bending during final thinning. Figure 3.15 Examples of artefacts during FIB-SEM acquisition of NMC811: (a) Curtaining artefacts caused by ion channelling during milling, resulting in vertical striations; (b) Shine- through artefact, where material from deeper layers is imaged due to electrons escaping through an open pore; and (c) Sample charging visible as bright regions. Scale bar represents 2 μm. In FIB-SEM tomography, curtaining artefacts are typically more pronounced due to the higher ion beam currents required for milling large volumes quickly, much higher than those used in lamella polishing. To mitigate this, a dense platinum layer was first deposited to achieve a more uniform sputtering rate, followed by the application of the Stripe Filter plug-in in ImageJ178. This plug-in utilises a combined wavelet-Fourier filtering approach, integrating wavelet analysis for spatial localisation and Fourier Transform (FFT) for frequency isolation. Daubechies 5 (db10) wavelets at decomposition level 3 and 𝜎 = 8 were selected for their 62 3.6 Artefact Mitigation Strategies effectiveness in reducing high-frequency noise while maintaining image details. The FFT was applied to the vertical detail coefficients to isolate and suppress the curtain artefacts, after which inverse transformations were performed to restore the image. The resulting images, shown in Figure 3.16, demonstrated the superior performance of the combined wavelet-FFT method in effectively removing the artefacts without significant loss of image quality. Figure 3.16 Comparison of curtain artifact removal techniques in SC2 FIB-SEM images. (a) shows the original image with visible curtain artifacts. The vertical detail coefficients isolated through the filtering process and the resulting image after applying the filter are as follows: (b) and (c) for wavelet-only filtering, (d) and (e) for FFT-only filtering, and (f) and (g) for the combined wavelet-FFT approach. Scalebar represents 2𝜇m. Shine-through artefacts (STA) occur when large, open pores allow electrons to escape through the sample, causing deeper material layers to be imaged across multiple consecutive slices. This effect introduces angled streaks in the reconstructed volume, particularly in the Y-Z projection, and can result in the misclassification of large pores as solid material. As illustrated in Figure 3.17, STA originates from electron penetration through open pores, leading to repeated imaging of the same feature across multiple slices. To address this, datasets were transposed into the orthogonal Y-Z plane using Dragonfly software to enhance visual detection of these artefacts. The STA-affected regions were then accurately segmented and removed using a 3D U-Net model, effectively distinguishing them from true pore structures. 3.7 Summary 63 Charging artefacts, seen as bright patches resulting from charge accumulation, particularly on the carbon binder, were mitigated by imaging at low voltage (2 kV) and low current (300 pA), and by using a backscattered electron detector to reduce susceptibility to surface charging. Figure 3.17: (a) Schematic representation of the FIB-SEM tomography setup with sample stage tilted to 52°, positioning the ion beam perpendicular to the sample surface for uniform milling (b-d) origin of the shine-through artefact (STA), where electrons escape through large open pores, imaging deeper layers across multiple slices. In (c) slice n, the STA-affected region is highlighted; in (d), slice n+10 (approximately 500 nm deeper), (e) angled streaks in the Y-Z projection that arise because the artefact remains visible in nearly the same location across multiple slices. Figure produced by Morzy29 3.7 Summary This chapter has outlined the experimental methods and analytical techniques employed to investigate the structural and electronic evolution of NMC811 cathode materials. The research strategy integrates advanced electron microscopy and spectroscopy approaches to establish direct correlations between electrochemical cycling, microstructural degradation, and changes in electronic structure at multiple length scales. Electron Energy Loss Spectroscopy (EELS) served as the primary analytical technique, offering unparalleled sensitivity to electronic structure changes in transition metal oxides. The EELS methodology was carefully optimised through theoretical analysis of multiple scattering effects, establishing an ideal specimen thickness regime ( 𝑡/𝜆 ≈ 0.5) that balances signal strength and spectral quality. 64 3.7 Summary To achieve this optimal thickness, a significant methodological advancement was developed in the form of a modified FIB lamella mounting approach. This innovation enabled the reliable preparation of ultrathin specimens (~ 40 - 65 nm) of mechanically challenging NMC811 materials, critical for acquiring high-quality EELS data with minimal plural scattering artefacts. Complementary techniques including high-resolution STEM imaging and 3D FIB-SEM tomography provided contextual microstructural information, allowing degradation processes to be characterised across multiple length scales. Together, these methods formed a powerful analytical framework for investigating the complex interplay between structural evolution and electronic changes in NMC811 cathodes during electrochemical cycling, the results of which will be presented in the following chapters. Chapter 4 65 Chapter 4 Microstructural Analysis and Tortuosity Characterisation of Calendered NMC811 Electrodes This chapter establishes a multi-method framework for quantifying microstructure-transport relationships in calendered NMC811 electrodes. A 3D U-Net segmentation approach differentiates active material, carbon-binder domain, and pore network whilst accounting for shine-through artefacts, enabling directional tortuosity analysis that reveals calendering-induced anisotropy obscured by conventional bulk measurements. Unlike most studies that focus exclusively on pore-phase transport, this work quantifies both ionic and electronic tortuosity in parallel, demonstrating that optimal performance requires concurrent optimisation of both pathways. Three fundamentally different tortuosity methods are compared: Dragonfly (graph-based), TauFactor (finite-difference), and symmetrical cell EIS (performed by Dr Kumar Raju). For single-crystal (SC) electrodes, convergent trends across all methods validate that microscale improvements translate to macroscopic performance. For polycrystalline (PC) electrodes, method divergence reveals scale- dependent behaviour where severe local cracking is partially compensated by alternative pathways at larger length scales. SC electrodes benefit from heavy calendering to 25 % porosity, achieving anisotropy ratios below 1 in both phases with superior rate capability (>85 % capacity retention at 2 C), enabled by their smaller particle size and mechanical resilience. PC electrodes require conservative calendering to 35 % porosity, as their larger particle size and weak grain boundaries cause extensive intergranular cracking that increases ionic resistance by 25 % despite improving electronic contact. The complementary strengths of imaging-based methods and electrochemical validation are essential for complete characterisation; future work with replicate measurements and plasma FIB technology would strengthen statistical confidence for industrial design guidelines. 66 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals Acknowledgement: The NMC811 electrodes used in this study were prepared and electrochemically tested (half-cell EIS, galvanostatic intermittent titration technique, hybrid pulse power characterisation, and long-term cycling) by Dr Kumar Raju, and have been published in our recent work 114 The work presented in this chapter builds upon that study and includes an extension involving tortuosity analysis. Dr Kumar Raju also assisted with running symmetrical cell EIS measurements under blocking conditions, which I subsequently analysed as part of this work. I performed all the FIB-SEM tomography, image analysis, implementation of the computational models using Dragonfly and TauFactor, and the comprehensive correlation of microstructural characteristics with electrochemical performance data. Dr Steve Kench provided advice on the error analysis for TauFactor. 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals NMC811, a widely adopted cathode material in lithium-ion batteries (LIBs), combines high energy density with acceptable cycling stability, making it central to next-generation energy storage systems. 179,180 As discussed in Section 2.1, LIB performance depends on two parallel transport networks: ionic transport through the electrolyte-filled pore phase and electronic transport through the solid phase. The composite electrodes consist of three primary phases: the active material (NMC811) which serves as the lithium storage medium; the carbon-binder domain (CBD), comprising conductive carbon additives and polymer binder, provides mechanical cohesion and electronic connectivity; and the pore phase, filled with electrolyte, which supports ionic transport.181,182 The microstructural arrangement of these phases, directly governs the efficiency of both transport pathways and ultimately determines electrode performance. Pores within the electrode structure serve dual roles that fundamentally affect both mass and electron transport.183 For Li-ion diffusion, pores act as pathways for electrolyte penetration, with high porosity increasing the number of available diffusion routes and enabling faster Li- ion supply to the electrode.184–186 However, from an electron transport perspective, pores can interfere with electron movement because electrons travel via direct contact between electrode materials. To attract Li ions inside the electrode, electrical transport must be enhanced by increasing contact between electrode materials, because electrochemical reactions require both ionic transport through the electrolyte and electronic conduction through the solid network to access reaction sites throughout the electrode.183 Pores therefore structurally and fundamentally affect Li+ transport through the interaction of mass and electron transport.183,187 Consequently, they must be designed considering the interplay between these two transport mechanisms.187,188 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals 67 As discussed in Section 2.4.4, transport limitations in either the ionic or electronic pathway can generate concentration gradients and non-uniform utilisation through the electrode thickness, which promotes spatially heterogeneous degradation. 4.1.1 Calendering and Microstructural Effects on Transport To achieve high volumetric energy densities, electrodes require calendering, a standard industrial process where electrodes are passed through heated rollers to compress the electrode and pack more active material per unit volume, as illustrated in Figure 4.1.189,190 Calendering is a critical processing step for optimising electrode performance. It improves electrical conductivity by enhancing contact between electrode components, particularly strengthening the electronic percolation network formed by carbon additives and improving particle-to-particle contact.191,192 This enhanced interfacial contact prevents the disconnection of carbon and active particles, reduces polarisation, and improves rate capability and cycle life. The improved particle connectivity also contributes to better cycling stability by reducing resistive losses over time.190,193–195 However, the response to calendering differs fundamentally between polycrystalline (PC) and single-crystal (SC) particle morphologies, as illustrated in Figure 4.1. However, whilst calendering optimises electronic conductivity through improved solid-phase contact, it simultaneously reduces pore volume and size, which restricts ionic transport pathways through the electrolyte-filled pore network. 196,197 An excessively strong calendering process can cause particles to break or reduce electrolyte penetration into the electrode.198,199 The calendering density must therefore be carefully selected based on the power and energy requirements of the application. Excessive calendering can lead to cracking of cathode particles, increasing the active material's surface area and accelerating degradation.200,201 Ni- rich cathodes such as NMC811 are particularly vulnerable to surface-driven degradation processes,17,202,203 necessitating further investigations to determine optimal calendering densities for these advanced cathodes. 68 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals Figure 4.1: Schematic illustration of the calendering process and its microstructural effects on polycrystalline (PC) and single-crystal (SC) NMC811 particles. Calendering compresses the electrode using opposing rollers, applying a force (FN) that enhances interparticle contact. For PC particles (top row), the applied pressure can induce intergranular cracking due to the presence of grain boundaries. In contrast, SC particles (bottom row), lacking internal boundaries, undergo deformation while largely maintaining structural integrity. The light blue outlines represent the carbon-binder domain (CBD), which surrounds the active material particles and facilitates electronic conductivity and structural cohesion. The interplay between the active material, CBD and pore phase plays a decisive role in electrode performance.196 Calendering modifies their arrangement and connectivity, introducing anisotropy that alters transport pathways in the through-plane and in-plane directions. Figure 4.2 summarises the trade-off: high-porosity electrodes exhibit poorer particle contact and higher contact resistance (𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡), alongside longer ionic pathways and higher ionic resistance ( 𝑅𝑖𝑜𝑛 ). In low-porosity calendered electrodes, improved particle contact and shortened ionic pathways reduce 𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡 and 𝑅𝑖𝑜𝑛 , respectively. Optimising electrode performance therefore requires balancing electronic and ionic transport. 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals 69 Figure 4.2: Schematic of Li-ion and electron transport in high-porosity (left) and low-porosity calendered (right) electrodes. Grey spheres: active material; black circles: carbon additives; purple regions: electrolyte-filled pores; green dashed lines: electronic percolation pathways; red regions: contact resistance (𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡). Adapted from Kim et al.204 4.1.2 Tortuosity: Definitions, Measurement, and Empirical Relations A key parameter for evaluating transport limitations is tortuosity (𝜏), which describes how much the effective transport path in a porous or percolating medium deviates from a straight line.205 In composite battery electrodes, tortuosity is relevant for two networks: ionic tortuosity (𝜏𝑖𝑜𝑛𝑖𝑐 ) characterises transport through the electrolyte-filled pore space, while solid-phase (electronic) tortuosity ( 𝜏𝑒𝑙𝑒𝑐𝑡𝑟𝑜𝑛𝑖𝑐 ) describes the geometric complexity of the percolating electron-conducting network formed by the active material and carbon–binder domain. The active material provides the structural framework that shapes the pore architecture, the CBD contributes electronic connectivity while occupying pore volume, and the remaining pore space defines the pathways available for ionic transport. Calendering modifies these phases concurrently: it increases particle contact and can improve solid-phase connectivity, while compressing and reorienting the pore network, which can increase τ𝑖𝑜𝑛𝑖𝑐 depending on how pore connectivity and alignment evolve.206 As a result, tortuosity influences effective transport, concentration polarisation and rate capability, with consequences for capacity retention and overall efficiency.189,207 Kim et al.204 investigated the effect of calendering on NMC532 electrodes with different porosities (50 %, 40 %, 30 %, and 20 %) using EIS analysis with blocking symmetric cells. Their findings revealed that whilst the 30 % porosity electrode exhibited the smallest 𝑅𝑖𝑜𝑛 (14.8 𝛺 𝑐𝑚), indicating optimal ionic transport pathways, the 20 % porosity electrode demonstrated the best rate capability due to its smallest 𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡 (2.8 𝛺 𝑐𝑚² ), which enhanced electronic conductivity. This demonstrates that overall electrochemical performance is governed by the combined effect of both resistances rather than either parameter alone. Importantly, their work demonstrated that rate performance reflects coupled limitations in ionic and electronic transport, rather than being determined solely by ionic pathway characteristics. These insights 70 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals underscore the importance of characterising both transport networks separately to understand how processing conditions like calendering affect overall electrode performance. To quantify how microstructural characteristics affect ionic transport processes, researchers typically connect ionic tortuosity to measurable electrode properties through established relationships. One of the most important of these is the MacMullin number ( 𝑁𝑀 ), a dimensionless parameter that quantifies the ratio of bulk electrolyte conductivity to effective conductivity within a porous medium. This number provides a practical measure of ionic transport limitation and serves as a bridge between easily measured properties and the more complex ionic tortuosity:208 𝑁𝑀 = 𝜅 𝜅𝑒𝑓𝑓 = 𝜏 𝜀 Equation 6 Where 𝜅 is the ionic conductivity of the electrolyte solution, 𝜅𝑒𝑓𝑓 is the effective ionic conductivity within the porous electrode, 𝜏 is tortuosity, and 𝜀 is porosity. The MacMullin number forms the basis for empirical laws linking microstructure to macroscopic conservation laws.208 The Bruggeman relation is frequently used in battery research to describe the relationship between porosity and tortuosity:209 𝜏 = 𝜀1−𝛼 Equation 7 For spherical particles, the Bruggeman exponent, 𝛼, is typically 0.5 but experimental results often deviate from this value.34,157 Researchers have modified the Bruggeman equation by adjusting 𝛼 and introducing a scaling factor, 𝛾, to better fit experimental data. Thorat et al.35 extended the equation to: 𝜏 = 𝛾 𝜀1−𝛼 Equation 8 with a scaling parameter 𝛾 of 1.8 to align predictions with AC impedance spectroscopy results for battery separators and cathode materials. Zacharias et al.32 refined the model by linking 𝛼 and 𝛾 to electrode compositions. They observed higher 𝛾 values (2.5 and 2.6) and lower 𝛼 values (1.27 and 1.28). However, derived values for 𝛼 and 𝛾 vary significantly across different materials and microstructures due to differences in manufacturing, composition, and pore size distribution. Some extrapolations even produce tortuosity values below unity, contradicting the physical definition of 𝜏. These 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals 71 findings cast doubt on the usefulness of this method. As a consequence, the application and interpretation of 𝛼 and 𝛾 values have to be analysed with caution. Studies have reported conflicting results regarding the validity of the Bruggeman relation when compared to tomography-based calculations.210,211 While the correlation aligns well with spherical structures, it is less reliable for connected solid phases and complex porous networks. Thus, porosity-tortuosity relationships are valid only for homogeneous microstructures resembling the original assumptions of the Bruggeman model. Examining tortuosity in both the pore and active material phases provides a more comprehensive understanding of how structural changes, such as those induced by calendering, affect electrochemical performance. While optimising microstructure can improve transport pathways and enhance performance, increased tortuosity can hinder ion transport and exacerbate resistive losses, making tortuosity analysis an essential tool for guiding electrode design and fabrication. 4.1.3 Characterisation Techniques: X-ray CT versus FIB-SEM Traditional imaging techniques such as X-ray computed tomography (X-ray CT) are widely used to study LIB electrode microstructures, but they can be limited for resolving the carbon- binder domain (CBD) and pores in composite cathodes.181,182 In many practical datasets, X-ray CT is constrained by spatial resolution and, more importantly, by weak phase contrast between the CBD and the pore phase due to their similar X-ray attenuation.181,182 This can hinder reliable segmentation, introducing uncertainty into derived microstructural metrics such as porosity, pore connectivity and tortuosity. Distinguishing these phases is important because they support different transport pathways: the CBD primarily contributes to electronic conduction, whereas pores provide ionic transport through the electrolyte. If the CBD and pores are combined into a single phase during segmentation, the resulting tortuosity values are difficult to interpret and may not represent either pathway accurately. Although X- ray CT offers a larger field of view than FIB-SEM, these segmentation limitations can restrict its utility for transport analysis when CBD and pore discrimination is required. Focused ion beam scanning electron microscopy (FIB-SEM) addresses these limitations by providing nanometre-scale resolution and strong phase contrast in backscattered electron imaging, enabling robust separation of active material, CBD and pores.212 This phase-resolved 3D reconstruction allows ionic transport to be analysed through the pore network and electronic transport through the percolating solid phase (active material + CBD), enabling transport metrics that are difficult to obtain reliably from X-ray CT when CBD-pore contrast is poor.213 Importantly, this separation supports a mechanistic assessment of how calendering modifies both pathways. Kim et al.'s work204 showed that optimal electrode performance 72 4.1 Introduction: NMC811 Cathodes and Transport Fundamentals requires balancing𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡 (linked to electronic transport limitations) and 𝑅𝑖𝑜𝑛(linked to ionic transport limitations), rather than minimising either resistance in isolation. FIB-SEM is therefore particularly well suited to analysing calendering-induced changes in pore alignment, CBD distribution and particle connectivity that jointly govern ionic and electronic transport. The main limitation of FIB-SEM is field of view where typical reconstructed volumes are tens of micrometres rather than millimetres, as obtained by X-ray CT. For this methodology-driven study, the trade-off is justified because the higher resolution and phase contrast are required to develop reliable segmentation workflows and to validate computational tortuosity estimates against symmetric-cell electrochemical impedance spectroscopy. Once established at the FIB-SEM scale, the approach can be extended using plasma FIB systems, which can access fields of view exceeding 100 𝜇𝑚 while maintaining nanoscale resolution, enabling larger and more statistically robust studies.33 4.1.4 Chapter Aims and Structure This chapter investigates the relationship between calendering, microstructure, and tortuosity in NMC811 electrodes, with the following aims: 1. How do different methodologies, Dragonfly, TauFactor, and EIS, compare in their ability to quantify tortuosity, and can their complementary strengths be integrated for a comprehensive analysis? (Addressed in Chapters 4.2.4 to 4.2.6 for the methodology explanation, and analysed in Chapters 4.3.2 to 4.3.6) 2. What are the effects of calendering on microstructure, directional transport properties, anisotropy and electrochemical performance in single-crystal (SC) and polycrystalline (PC) NMC811 electrodes? (Addressed in Chapter 4.3.6 to 4.3.7) 3. How do calendering-induced changes impact tortuosity-porosity relationships, particularly deviations from the Bruggeman model, and what are the implications for long-term cycling stability and battery design? (Addressed in Chapter 4.3.5 and 4.3.8) To address these questions, a multimodal workflow was implemented, combining electrochemical cycling, FIB-SEM imaging, segmentation, and transport analysis using multiple tools (Figure 4.3). This approach enables detailed, phase-resolved insights into how calendering alters microstructural features and transport pathways. 4.2 Methodology for Microstructure and Transport Analysis 73 Figure 4.3: Schematic overview of the experimental workflow for investigating the effects of calendering on NMC811 cathode microstructure and transport properties. (1) Electrodes are calendered to different densities using heated rollers. (2) Samples are electrochemically cycled to simulate operational conditions. (3) FIB-SEM tomography is performed to acquire high- resolution 3D datasets. (4) The tomograms are segmented into distinct phases—active material, binder (CBD), and pore. (5) Tortuosity and transport pathways are quantified via graph-based analysis to evaluate directional anisotropy and the impact of microstructural changes. 4.2 Methodology for Microstructure and Transport Analysis This section outlines the methodologies employed in this study, starting with electrode preparation processes detailed in Section 4.2.1, where the fabrication of both single crystal (SC) and polycrystalline (PC) NMC811 electrodes is described. This is followed by the electrochemical cell assembly and protocols, discussed in Section 4.2.2 which cover the configurations of coin cells, three-electrode cells, and pouch cells, along with the associated testing procedures, including formation cycles, rate capability, and HPPC tests. The cells were assembled and tested by a collaborator, Dr Kumar Raju. 74 4.2 Methodology for Microstructure and Transport Analysis The subsequent sections focus on microstructure analysis. Section 4.2.3 delves into segmentation and 3D reconstruction using supervised machine learning, specifically a 3D U- Net convolutional neural network, to isolate distinct phases within the FIB-SEM datasets, such as active material, pore space, and carbon-binder matrix. Finally, the theoretical background and practical application of tortuosity analysis are discussed across three methodologies. The first approach, detailed in Section 4.2.4, involves using Dragonfly software to create sparse and dense graphs for tortuosity calculations. The second method, outlined in Section 4.2.5, leverages the Python-based TauFactor package to calculate tortuosity using voxel-based datasets. The third technique, presented in Section 4.2.6, employs symmetrical cell Electrochemical Impedance Spectroscopy (EIS) to determine ionic resistance and tortuosity through Nyquist plot analysis. Together, these methodologies provide a comprehensive framework for evaluating the impact of calendering protocols on the microstructural and transport properties of NMC811 electrodes. 4.2.1 Electrode fabrication Electrode fabrication and electrochemical characterisation (Section 4.2.2) were carried out by Dr Kumar Raju; full experimental details are provided in the published literature.75 The methods outlined below provide an overview to give readers context for the subsequent microscopy analysis presented in this thesis. Single-crystal (SC) and polycrystalline (PC) NMC811 powders were obtained from Sinochem and MTI Corporation, respectively, and their morphologies were confirmed by XRD and SEM characterisation as reported in recent literature. 214,215 The SC and PC powders were dried under vacuum overnight before slurry preparation. Slurries were formulated with a 90:5:5 weight ratio of NMC811, PVDF binder, and Super P conductive carbon, dispersed in N-methyl-2-pyrrolidone (NMP). Films were cast onto 16 μm aluminium foil and dried at 120 °C overnight, achieving ~100 μm initial thickness. Calendering was conducted using a two-roller press at 80 °C, with both forward and reverse passes applied to enhance particle contact. To achieve industrially relevant density, both the SC and PC electrodes were calendered to 35 % and 25 % porosity. The SC electrodes have 2.7, 3.5, and 4.2 g cm−3 densities for uncalendered, and calendered to 35% and 25% porosity, respectively. The PC electrodes have 2.23, 3.13, and 3.58 g cm−3 densities for uncalendered, and calendered to 35% and 25% porosity, respectively. 4.2 Methodology for Microstructure and Transport Analysis 75 Porosity (𝜀) was calculated based on mass fractions and intrinsic densities of each component: 𝜀 = 𝑀 (𝑋𝑁𝑀𝐶/𝛿𝑁𝑀𝐶 + 𝑋𝑃𝑉𝐷𝐹/𝛿𝑃𝑉𝐷𝐹 + 𝑋𝐶/𝛿𝐶) 𝑉 Equation 9 where M and V represent the mass and volume of the electrodes, while X and δ denote the mass fraction and density of the NMC, PVDF, and carbon black components within the electrode. 4.2.2 Electrochemical cell assembly and protocols The cells were assembled in an Argon-filled glovebox to minimise contamination by moisture and oxygen. Two main configurations were prepared: 2032-type coin cells and single-layer pouch cells. The coin cells included half-cells with a 13 mm cathode, a 15 mm lithium metal counter electrode (Hohsen), and a 19 mm Celgard separator soaked in 42 𝜇𝑙 of LP57 electrolyte (SoulBrain). Full cells were similarly constructed with a 14 mm cathode, a 15 mm graphite anode, and a 260 𝜇𝑚 thick GF/B grade glass fiber separator saturated with 100 μl of electrolyte comprising 1.3 M LiPF6 in EC:EMC:DEC (3:5:2 by volume) with additives including 9:1 wt % FEC, 0.5 wt % VC, and 0.2 wt % LiBF4 (E-lyte Germany). After assembly, the half-cells underwent three formation cycles using a constant-current constant-voltage (CCCV) protocol at C/20 between 3-4.3 V. Full cells were tap-charged to 1.5 V, rested for 10 hours, and subjected to three formation cycles at C/20. Post-formation, rate capability and cycle stability tests were conducted to evaluate performance. In addition to the coin cells, three-electrode PAT cells were configured with an 18 mm cathode and anode separated by a 260 𝜇m thick GF/B grade glass fiber separator soaked with 100 𝜇l of LP57 electrolyte (1 M LiPF6 in EC/EMC 3:7 vol %). A lithium metal ring electrode encased in an insulating sleeve acted as the reference electrode. These cells were charged galvanostatically at C/20 to 3.05 V (vs LTO) using a Biologic VMP3 potentiostat, held at this voltage for 60 hours with current monitoring, and subsequently discharged at C/20 to 1.45 V. Single-layer pouch cells were also assembled with heavily calendered electrodes paired with graphite anodes. A Celgard separator, soaked in 0.5 ml of E-lyte electrolyte, was used. After the formation cycles, these cells were charged to 4.3 V at C/2 to prepare for Hybrid Pulse Power Characterisation (HPPC) testing. All electrochemical measurements were conducted in a climate chamber maintained at 26 °C. Electrochemical Impedance Spectroscopy (EIS) was performed using a Biologic BCS 805 Series potentiostat, with a scanning frequency range of 10 kHz to 10 mHz and an applied amplitude of 10 mV at a cell potential of 3.8 V. For the HPPC test, the pouch cells were charged to 4.3 V, rested for one hour, and discharged to 10 % depth of discharge (DOD) at C/2. A 10-second 3C 76 4.2 Methodology for Microstructure and Transport Analysis discharge pulse followed by a C/2 charge pulse with 40 seconds of rest time was repeated at 10% DOD intervals. Intermediate discharges were performed at C/2 to measure potential- dependent impedance, and the test was repeated before cycling and after every 100 cycles. Area-specific impedance (ASI) values were calculated based on changes in cell voltage during current pulses, using the geometric area of the cathode. These calculations provided key insights into the impedance behaviour and its evolution during cycling, contributing to the understanding of ionic resistance and overall electrode performance.216 4.2.3 Segmentation and 3D Reconstruction of FIB-SEM Images FIB-SEM tomography datasets were subjected to a systematic pre-processing workflow to improve image quality and minimise artefacts that could compromise segmentation and subsequent quantification. Regions of interest (ROIs) were first cropped to exclude unnecessary features, such as the milled pillar edges, protective Pt and C layers, and redeposited material at the trench base. To correct for the vertical compression introduced by the 28 ° angle between ion and electron beams, the images were geometrically stretched along the y-axis. Following curtaining artefact suppression, as explained in Section 3.6, image alignment was performed using the sum of squared differences (SSD) method in Dragonfly to correct for drift and inter-slice misalignments introduced during acquisition. The SSD algorithm operates by minimising the cumulative squared intensity differences between corresponding pixels of adjacent slices, effectively identifying the optimal lateral shift required to maximise structural continuity across the dataset. This alignment procedure was critical for ensuring that features such as particle edges, pores, and cracks remained spatially coherent in the reconstructed volume. Accurate alignment was essential not only for visual interpretation but also for downstream quantitative analyses such as segmentation, graph construction, and tortuosity calculation. To address the shine-through artefacts (STA), pre-processed datasets were imported into Dragonfly software and transposed to enhance the visibility of STA as angled streaks in the new Y-Z projection. A 3D U-Net convolutional neural network was then employed to classify the data into four categories: NMC811 particles, the carbon-binder matrix, small pores, and STA, as shown in Figure 4.4. Total pore segmentation was achieved by combining the small pores and STA categories, enabling detailed analysis of ion transport. Similarly, the active material phase was derived by combining the carbon-binder matrix with NMC811 particles, facilitating the study of electron transport within the reconstructed structure. 4.2 Methodology for Microstructure and Transport Analysis 77 Figure 4.4: Example of 3D-UNet based segmentation of FIB-SEM tomography data based on secondary electron detector images. The SEM images were reconstructed in the Y-Z plane (perpendicular to the FIB slicing plane, current collector is towards the bottom of the image, separator side towards the top). Overlay of the segmented regions of (a) shine-through artefacts (STA), (b) carbon binder matrix, (c) pore, (d) active material. The model was trained on 100 manually segmented slices drawn from six FIB-SEM datasets, with pixel-wise labelling into the four defined classes. A patch size of 32 × 32 × 32 voxels was used during training, with the Dice coefficient as the loss function to optimise segmentation accuracy. The Adadelta optimisation algorithm was adopted to ensure stable convergence. This deep-learning-based approach effectively addressed artefacts such as STA and enabled accurate, consistent segmentation across datasets. Each slice of the FIB-SEM image was taken at 60 nm intervals, with a cross-section pixel size of 10×10 nm2. This voxel geometry introduces significant anisotropy, making measurements in the slicing (Z) direction less reliable. Therefore, all measurements in this chapter are taken in the through-plane direction (perpendicular to the current collector) or the in-plane direction (parallel to the current collector), as illustrated in Figure 4.5. 78 4.2 Methodology for Microstructure and Transport Analysis 4.2.4 Tortuosity analysis (Dragonfly) To quantify the complexity of transport pathways in the segmented 3D datasets, tortuosity was calculated using skeletonised graphs in Dragonfly. Both sparse and dense graphs were generated from the segmented pore networks, representing simplified and complete connectivity respectively. Tortuosity was then computed by identifying two regions of interest (ROIs) within each dataset, placed to probe transport along the X (in-plane) and Y (through-plane/calendering) directions, as shown in Figure 4.5a. The Z-axis corresponds to the direction of serial sectioning during FIB-SEM acquisition. Figure 4.5: (a) The reconstructed volume of calendered single crystal electrode (SC-25) from FIB-SEM imaging. The blue arrow corresponds to the through-plane direction, which is the direction of calendering, perpendicular to the current collector, red arrow corresponds to in- plane direction, parallel to the current collector and green arrow represents the direction of serial sectioning during FIB-SEM acquisition. (b) Schematic comparison of sparse graph path (yellow), dense graph path (red) and Euclidean path (purple). The sparse graph path simplifies connectivity with fewer nodes and edges, offering computational efficiency, while the dense graph path captures all possible connections for detailed structural analysis. The Euclidean path represents the straight-line distance between the start and end points. Dijkstra’s algorithm was used to determine the shortest path between the source and destination ROIs. For each path, the algorithm calculates the total travelled distance across the network of edges, and the tortuosity (τgeo) is expressed as the ratio between the path length (Δ𝑙) and the straight-line (Euclidean) distance (Δx) between the ROIs:151 𝜏𝑔𝑒𝑜 = 𝛥𝑙 𝛥𝑥 Equation 10 4.2 Methodology for Microstructure and Transport Analysis 79 In Dragonfly, the sparse graph and dense graph differ significantly in their representation of connectivity. Sparse graphs simplify connectivity by reducing nodes and edges, improving computational efficiency and visual clarity, especially for large datasets. In contrast, dense graphs retain all possible connections between nodes, capturing subtle structural details but requiring significantly greater memory and processing power, which can exceed hardware limitations during 3D rendering. The choice between sparse and dense graphs represents a trade-off between computational efficiency and analytical precision. Figure 4.5b contrasts the two approaches, and Figure 4.6 illustrates the overlaid pore networks on reconstructed active material structures for SC and PC electrodes. Figure 4.6: Active material in grey with overlaid pore network visualisation (white nodes with multi-coloured edges representing the difference in lengths), demonstrating the complex transport pathways through the structure of PC in (a), (b) and (d), and SC in (b). 80 4.2 Methodology for Microstructure and Transport Analysis To account for geometric bottlenecks, a throat-weighted approach was also implemented. In this case, edge weights were derived from the inverse square of the cross-sectional area of the throat associated with each connection, such that narrower throats, which present higher resistance, received higher weights: 𝑤𝑡ℎ𝑟𝑜𝑎𝑡 = 𝑚𝑎𝑥(𝐴𝑡ℎ𝑟𝑜𝑎𝑡)2 𝐴𝑡ℎ𝑟𝑜𝑎𝑡 2 Equation 11 where 𝐴𝑡ℎ𝑟𝑜𝑎𝑡 is the cross-sectional area of the throat associated with the edge, and max(𝐴𝑡ℎ𝑟𝑜𝑎𝑡) is the maximum throat area in the dataset. These weights are then used in Dijkstra’s algorithm to calculate the shortest path, accounting for transport resistance. Weighted graphs enable Dijkstra’s algorithm to factor in physical resistance, offering a more realistic representation of ionic transport pathways, but require higher computational power. In non-weighted graphs, all edges are considered equal, and are useful when bottleneck effects are negligible. Dijkstra’s algorithm, first published in 1959,217 iteratively identifies the shortest cumulative- weighted path between nodes. In weighted graphs, the shortest path minimises the sum of resistive weights, while in non-weighted graphs it minimises the number of steps. Figure 4.7 illustrates the distinction between the two approaches. Dragonfly's graph tortuosity analysis calculates tortuosity for multiple pathways between the user-defined input and output ROI boxes. These boxes are positioned to probe transport in specific directions: through-plane (blue boxes) and in-plane (red boxes), as illustrated in Figure 4.5a. For each electrode sample and direction, Dijkstra’s algorithm was used to compute shortest paths for many start-end node pairs between the source and destination ROI boxes, producing a distribution of path lengths through the graph network. The tortuosity, 𝜏 is computed for each individual path according to Equation 10. The error bars presented in all Dragonfly-based tortuosity results represent the standard deviation of the distribution of tortuosity values across all paths between the input and output boxes. This standard deviation captures the variability in pathway complexity within the pore network, reflecting both geometric heterogeneity and the presence of multiple transport routes with differing levels of tortuosity. 4.2 Methodology for Microstructure and Transport Analysis 81 Figure 4.7: Comparison of weighted (left) and non-weighted (right) graphs illustrating the shortest path calculation using Dijkstra’s algorithm. In the weighted graph, edge weights influence the shortest path selection, leading to a route that minimises the total weight (red edges). The non-weighted graph treats all edges as equal, resulting in a shortest path based solely on the number edges. 4.2.5 Tortuosity analysis (Taufactor) The segmented pore phase from Dragonfly was exported into TauFactor2,218 a GPU- accelerated Python tool that solves the steady-state diffusion equation for voxel-based datasets. The tortuosity factor (𝜏 ) was calculated in both the in-plane and through-plan direction. 𝜏 quantifies the reduction in effective transport caused by the geometry of the porous medium and is defined as the ratio of the effective transport coefficient (𝐷𝑒𝑓𝑓) to the intrinsic transport coefficient (𝐷) of the conductive phase, corrected for the volume fraction (𝜀) of the conductive region: 𝐷𝑒𝑓𝑓 = 𝐷𝜀 𝜏 Equation 12 TauFactor operates directly on the three-dimensional reconstructed electrode volume. FIB- SEM serial sectioning produced ~ 250 consecutive images at 60 nm slice spacing, with an in- plane pixel size of 10 × 10 nm2. After alignment and stacking, these slices form a voxelised dataset spanning ~15 μm in depth (250 slices × 60 nm), used for all tortuosity calculations. Individual 2D slices are shown in Figure 4.8a-b for visualisation. a. Throat-weighted b. Non throat-weighted 82 4.2 Methodology for Microstructure and Transport Analysis Transport was modelled by solving the steady-state diffusion equation throughout the reconstructed pore network: −𝛻 ⋅ (𝐷𝛻𝑐) = 0 Equation 13 Where 𝑐 is the concentration of the diffusing species, 𝛻 is the gradient operator, and 𝐷 is the diffusivity of the conductive phase. Dirichlet boundary conditions were applied to opposite faces of the volume to impose a fixed concentration gradient along the direction of interest (through-plane or in-plane). Boundary stability was improved using the ghost-node method, and additional over-relaxation iterations were performed after convergence to ensure numerical accuracy. For multi-phase systems, the effective diffusion coefficient (𝐷𝑒𝑓𝑓 ) was determined by the software using the relation between the intrinsic diffusion coefficients of the transport phase, 𝐷𝑝, and the respective volume fractions, 𝑓𝑝, of each phase: 𝐷𝑒𝑓𝑓 = ∑ 𝐷𝑝𝑓𝑝 𝑝 𝜏 = 𝐷𝑚𝑒𝑎𝑛 𝜏 Equation 14 where 𝐷𝑚𝑒𝑎𝑛 represents the volume-fraction-weighted average of the intrinsic transport coefficients (𝐷𝑝 ) of the active phases. This assumes straight transport paths through each phase in the applied direction. Since τ ≥ 1 by definition, it captures the increased resistance due to complex microstructure.219 To balance computational cost against representativeness, tortuosity was evaluated using two sampling strategies (Figure 4.8c-d), with error bars providing complementary information about the reliability and representativeness of calculated values. In the subvolume approach (Figure 4.8c), the reconstructed volume was partitioned laterally into four independent 5 × 5 µ𝑚 regions, each spanning the full 15 μm depth. Tortuosity was calculated independently for each subvolume using TauFactor's finite-difference solver, with error bars representing the standard deviation of these four measurements. This quantifies spatial heterogeneity of transport properties within the electrode microstructure: larger standard deviations indicate that transport characteristics vary significantly across different lateral positions, suggesting non-uniform microstructural features such as particle clustering, pore size gradients, or localised calendering effects, while smaller error bars indicate more homogeneous transport properties. 4.2 Methodology for Microstructure and Transport Analysis 83 Figure 4.8: Illustration of the 3D volume analysis approach for TauFactor tortuosity calculations. (a-b) Representative 2D binary images (single slices from the 3D FIB-SEM tomography datasets) showing segmented pore phase (white) for polycrystalline (PC) and single crystal (SC) electrodes. (c) 3D subvolume approach: The reconstructed 3D volume is divided into four 5 × 5 𝜇𝑚 subregions (numbered 1-4, shown in different colors), each analysed independently through TauFactor to calculate tortuosity across the entire 3D stack. (d) 3D full-volume approach: A larger 10 × 10 𝜇𝑚 depth domain is analysed as a single 3D dataset. The red dashed squares in panels (a-b) correspond to the 2D projection of the subvolume regions shown in 3D in panel (c), while the blue dashed squares correspond to the full-volume region in panel (d). In the full-volume approach (Figure 4.8d), a single 10 × 10 × 15 µ𝑚 domain was analysed, providing broader lateral sampling at higher computational expense. Uncertainty was assessed by calculating tortuosity across multiple sampling geometries of varying sizes (2 × 2 84 4.2 Methodology for Microstructure and Transport Analysis × 15 µ𝑚 , 5 × 5 × 15 µ𝑚 , and 10 × 10 × 15 µ𝑚 ) at different lateral positions within the reconstructed volume. The error bars represent the standard deviation across these configurations, capturing both the influence of sampling size and spatial variability, thereby indicating whether the chosen volume is sufficiently representative of the electrode's bulk transport behaviour. 4.2.6 Tortuosity analysis (EIS Symmetrical Cell) The symmetrical-cell assembly, electrochemical impedance spectroscopy (EIS) measurements, and impedance-spectrum fitting described in this section were all performed by Dr Kumar Raju. The author (May Ching Lai) used the fitted parameters, specifically the ionic resistance ( 𝑅𝑖𝑜𝑛 ), to calculate the tortuosity, Bruggeman exponent, and MacMullin number, and subsequently interpreted these results in the context of the microstructural analysis presented in this chapter. To experimentally determine tortuosity, symmetric cells were assembled using identical electrodes (uncalendered and calendered), with Celgard separators and 20 mM Tetrabutylammonium hexafluorophosphate (TBAPF6) in EC:EMC as the electrolyte. AC impedance spectroscopy was performed from 106 to 10-3 Hz with a 10 mV amplitude. The Nyquist plots were fitted using a transmission-line model (TLM) in EC-lab software to extract the 𝑅𝑖𝑜𝑛. Ion transport through porous structures is governed by the ionic resistance, within the electrolyte phase. This resistance directly influences charge transport in porous particle networks or structures, such as lithium-ion battery electrodes or separators. By determining the ionic resistance, 𝑅𝑖𝑜𝑛, through a cross-sectional area (𝐴) of a material with porosity (𝜀) and thickness (𝑑), and knowing the conductivity of the electrolyte (𝜅), the MacMullin number (𝑁𝑀) and tortuosity (𝜏) can be calculated by rearranging Ohm’s law using Equation 1: 𝑁𝑀 = 𝜏 𝜀 = 𝑅𝑖𝑜𝑛 𝐴 𝜅 𝑑 Equation 15 In impedance-based approaches, blocking electrodes or blocking conditions are commonly employed to isolate ionic transport phenomena. Under these conditions, no charge transfer occurs across the solid-liquid interface, resulting in an ideally polarizable surface.157 Blocking conditions can be experimentally achieved by using salts, such as tetrabutylammonium perchlorate (TBACLO4), that do not participate in electrochemical reactions within the measurement’s potential range. However, surface roughness or inhomogeneous current distribution can deviate the behaviour of electrodes from being ideally polarisable.220 4.2 Methodology for Microstructure and Transport Analysis 85 Electrolyte-filled pores in lithium-ion battery electrodes are modelled using the Transmission- Line Model (TLM)157 The TLM in Figure 4.9 represents: 1. Electronic resistance in the solid phase through a series of ohmic resistors, 𝑅𝐸𝑙 2. Ionic resistance in the electrolyte phase through ohmic resistors, 𝑅𝑖𝑜𝑛 3. Charge transfer reactions at the solid-liquid interface, whether faradaic or capacitive, modelled by surface impedance elements, 𝑧𝑠 Figure 4.9: (a) General transmission-line model (TLM) equivalent circuit for a porous electrode, representing charge transport in the solid (grey region) and electrolyte (blue region) phases. The solid phase is characterised by electronic resistances, rEl , while the electrolyte phase includes ionic resistances, rion . The charge transfer process at the solid-liquid interface is modelled by surface impedance elements, zs , which account for faradaic or capacitive reactions. (b) Simplified transmission-line model for porous electrodes under blocking conditions and with negligible electronic resistance (rEl ≪ rion), including ionic resistances, rion, and constant-phase elements, qs for capacitive behaviour. (c) Schematic representation of the porous electrode microstructure, highlighting an ionic pathway within the electrolyte phase through a crack in the structure, representative of the transport pathways modelled in the TLM. Adapted from refs157,221 86 4.2 Methodology for Microstructure and Transport Analysis In well-designed lithium-ion battery electrodes, 𝑅𝐸𝑙 is typically negligible (𝜅 > 0.1 S cm-1) due to conductive carbon additives.222,223 In contrast, the ionic resistance in the electrolyte phase (𝜅 < 0.01 S cm-1) dominates. As a result, 𝑅𝐸𝑙 can often be omitted from the TLM. Under blocking conditions, where faradaic charge transfer reactions are absent, the surface impedance elements, 𝑧𝑠, exhibit (non-ideal) capacitive behaviour, which is modelled using constant- phase elements (CPEs) which represents a more realistic representation of the double layer capacitance at the electrode surface. CPEs are used to account for surface roughness.220 When 𝑅𝐸𝑙 ≪ 𝑅𝑖𝑜𝑛 and blocking conditions are applied, the general TLM simplifies to a configuration with only 𝑅𝑖𝑜𝑛 and CPEs (𝑞𝑠 ), as shown in Figure 4.9b. This configuration ensures that: (i) ionic flow remains confined to the electrolyte phase, and (ii) no ionic flow enters the current collector (pure electron conductor) and no electronic flow enters the electrolyte (pure ion conductor) The simplified TLM, models the analytical electrode impedance 𝑍𝐸𝑙 as follows157: 𝑍𝐸𝑙 = √𝑅𝑖𝑜𝑛𝑍𝑠 𝑐𝑜𝑡ℎ (√ 𝑅𝑖𝑜𝑛 𝑍𝑠 ) = √ 𝑅𝑖𝑜𝑛 𝑄𝑠(𝑖𝜔)𝛾 𝑐𝑜𝑡ℎ (√𝑄𝑠(𝑖𝜔)𝛾𝑅𝑖𝑜𝑛) Equation 16 where 𝑅𝑖𝑜𝑛 = ∑(𝑟𝑖𝑜𝑛) , 𝑄𝑠 = ∑(𝑞𝑠) , 𝑍𝑠 = ∑(𝑧𝑠) , 𝜔 is the angular frequency, and 𝛾 is the constant phase exponent. The 𝑅𝑖𝑜𝑛of electrodes can be extracted from impedance measurements by fitting Equation 17 to experimental data. Specialised software tools such as Aftermath facilitates this process by extrapolating the low- and high-frequency regions of Nyquist plots to the x-axis. The high- frequency resistance, 𝑅𝐻𝐹𝑅, is obtained from the intercept of the high-frequency branch, which contributes to the electronic resistance of the electrode, the contact resistances, and the ionic resistance of the separator. The low-frequency branch extrapolation provides the sum of 𝑅𝐻𝐹𝑅 and one-third of the 𝑅𝑖𝑜𝑛, as described by Ogihara et al224 and Liu et al.:225 𝑍𝐸𝑙|𝜔𝑙𝑜𝑤→∞ = 𝑅𝑖𝑜𝑛 3 + 𝑅𝐻𝐹𝑅 Equation 17 4.3 Results and Discussion 87 After determining 𝑅𝑖𝑜𝑛, the tortuosity, 𝜏, of the electrodes can be calculated by rearranging Equation 15: 𝜏 = 𝑅𝑖𝑜𝑛 𝐴 𝜅 𝜀 2𝑑 Equation 18 where the porosity ε of the electrodes can be determined from the areal weight and the thickness of the electrodes and the factor of 2 accounts for the symmetry of the setup, where the measured impedance represents the sum of the impedances of two identical electrodes. This approach provides a robust methodology for analysing ionic resistance and tortuosity in porous battery electrodes, enabling identification of transport bottlenecks and optimisation of ionic pathways. 4.2.7 Statistical analysis (EIS Symmetrical Cell) Due to equipment availability, one symmetric cell was measured for each electrode condition: SC uncalendered (SC-UC), SC 35% porosity (SC-35), SC 25% porosity (SC-25), PC uncalendered (PC-UC), PC 35% porosity (PC-35), and PC 25% porosity (PC-25), giving six measurements in total (one per condition). This limits statistical interpretation of the EIS- derived transport metrics. In the imaging-based approaches (Dragonfly and TauFactor), error bars quantify intra-sample variability within a reconstructed volume; in contrast, the EIS results cannot capture inter-sample variability arising from electrode fabrication, calendering uniformity, or electrolyte infiltration because no replicate electrodes were measured. Repeated impedance acquisitions on the same cell are typically reproducible, yielding stable spectra, but this repeatability does not substitute for independent repeats. Accordingly, the EIS-derived tortuosity, Bruggeman exponent, and MacMullin number are treated as macroscopic estimates used primarily to assess relative trends across calendering conditions. These trends are interpreted alongside, and checked for qualitative consistency with, the microscopy-based tortuosity analyses. Replicate symmetric-cell measurements would be required for full statistical validation and are therefore identified as future work. 4.3 Results and Discussion This section combines advanced imaging techniques, computational modelling, and electrochemical analysis to deliver a comprehensive understanding of the structural and transport properties of SC and PC NMC811 electrodes. 88 4.3 Results and Discussion The chapter begins with Section 4.3.1, which examines the microstructural differences between uncalendered and calendered SC and PC electrodes. Using FIB-SEM, this section underscores its essential role in characterising crack propagation and analysing surface and cross-sectional morphologies before and after electrochemical cycling. This analysis highlights how calendering impacts crack formation and distribution within the electrode architecture. Section 4.3.2 to 4.3.4 focus on the methodologies employed to calculate tortuosity in NMC811 electrodes. These chapters include: • A comparative analysis of dense versus sparse graphs to quantify transport pathways. • An evaluation of throat-weighted versus non-throat-weighted approaches, highlighting their importance in tortuosity modelling. • A comparison of TauFactor and Dragonfly software, assessing their respective strengths and limitations in tortuosity calculations. Section 4.3.5 explores the tortuosity-porosity relationships in NMC811 electrodes, critically evaluating deviations from the traditional Bruggeman relationship. It investigates the influence of calendering on ionic and electronic resistance, tortuosity, and the MacMullin number for both SC and PC electrodes. This analysis integrates symmetrical cell Electrochemical Impedance Spectroscopy (EIS) and imaging-based methods. Section 4.3.6 discusses the electrochemical implications of the observed microstructural changes, linking tortuosity metrics to performance outcomes. It provides a comprehensive comparison of the various tortuosity calculation methods, emphasising their practical relevance and limitations. Section 4.3.7 compares the three techniques, Dragonfly, TauFactor, and EIS, used to evaluate tortuosity in SC and PC electrodes, focusing on both the through-plane and in-plane directions to elucidate anisotropic tortuosity. By analysing their complementary strengths, the comparison highlights how these methodologies collectively enhance our understanding of transport phenomena in calendered NMC811 electrodes and their impact on directional transport pathways. Section 4.3.8 integrates the findings thus far to propose optimisation strategies tailored to SC and PC NMC811 electrodes. It provides actionable insights for improving ionic and electronic transport properties through targeted calendering protocols, balancing structural integrity and performance to achieve optimal battery functionality. This Results and Discussion section not only highlights the structural and electrochemical changes induced by calendering but also establishes a framework for integrating 4.3 Results and Discussion 89 complementary methods to enhance the accuracy of tortuosity calculations. By bridging imaging-based and electrochemical approaches, it advances our understanding of how manufacturing processes impact electrode performance and informs future strategies for electrode design. 4.3.1 Microstructure comparison between uncalendered and calendered single and polycrystalline NMC811 Figure 4.10 illustrates SEM cross-sections of SC and PC electrodes prepared at varying porosities. There were no visible cracks across all the SC and PC apart from PC-25, which had undergone the most calendering. The presence of cracks highlights mechanical deformation during calendering as a key contributor to crack formation. Similar trends were reported by Heenan et al.,226 where approximately one-third of particles were defective prior to operation due to fabrication-induced cracking. To further understand the evolution of these microstructural changes, the following sections will explore cycled cells and their electrochemical implications, which form the basis of this chapter. Figure 4.10: Cross-section images of uncalendered, calendered to 35% and 25% porosities of poly and single crystal NMC811 prior to any electrochemical cyclings. Scalebar for the top and bottow rows are 2 𝜇𝑚 and 5 𝜇𝑚 respectively. 90 4.3 Results and Discussion After 300 cycles at a 1C rate, SEM surface images in Figure 4.11 reveal significant differences in cracking behaviour. SC electrodes retain their structural integrity across porosity levels, whereas PC electrodes show visible cracks on the surface, highlighted by red arrows, which increase in density and severity with decreasing porosity. These findings emphasise the impact of calendering on surface microstructure, particularly in PCs. Calendering compresses the particles, increasing their density across both electrode types. For PC electrodes, uncalendered secondary particles maintain a more spherical shape, while calendered particles exhibit significant compaction and deformation. While surface observations provide valuable insights, cross-sectional imaging is critical to understand subsurface damage. Figure 4.11: SEM surface images of single-crystal (top row) and polycrystalline (bottom row) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The red arrows indicate cracks that can be seen from the surface of calendered PCs. The scalebar is 10 𝜇𝑚. Compared to uncycled cross-sections (Figure 4.10), cycled cells (Figure 4.12) exhibit intragranular cracks (red arrows) across all samples, even in uncalendered materials. Cracks are more prominent in PC cycled cells, with intergranular cracks present even in uncalendered PC samples but significantly more extensive in the calendered PC samples. Magnified images in Figure 4.13 reveal intragranular cracks in SC samples and intergranular cracks in PC samples, with higher cracking density observed at the electrode surface compared to the bulk. The layered NMC structure complicates intragranular crack quantification, causing an underestimate, as these cracks are visible only along the (003) planes227–229 and SEM imaging cannot resolve nanoscale cracks smaller than 2 nm. 4.3 Results and Discussion 91 Figure 4.12: Cross-sectional SEM images of single-crystal (SC) and polycrystalline (PC) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The top row shows SC samples, where the structure remains largely intact across all porosity levels. The bottom row depicts PC samples, illustrating progressive intergranular cracking and structural degradation as porosity decreases. Red arrows in the SC samples highlight intragranular cracks, while the PC samples demonstrate severe compaction and cracking, particularly for calendered samples, indicative of higher mechanical stress and calendering-induced damage. The scalebar is 2 𝜇𝑚. Intragranular cracking arises from inhomogeneous stresses within the NMC crystal, driven by factors such as: (1) expansion/contraction of neighbouring primary particles, (2) transition metal inhomogeneity, (3) mismatched expansion/contraction at rock-salt and layered phase interfaces, and (4) uneven Li distribution, leading to varying levels of expansion and contraction. 230–232 These stresses are exacerbated at high SOC, where NMC exhibits reduced mechanical strength and Young’s modulus, making it more prone to crack formation.233 92 4.3 Results and Discussion Figure 4.13: Magnified cross-sectional SEM images of single-crystal (SC) and polycrystalline (PC) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The top row shows SC samples with intragranular cracks. The bottom row illustrates PC samples, highlighting significant intergranular cracking that increases in density and severity with decreasing porosity. The scalebar for top and bottom rows are 1 𝜇𝑚 and 2 𝜇𝑚 respectively. To understand crack propagation in secondary particles, consecutive images obtained through FIB-SEM provide valuable insights. The sequence of images in Figure 4.14 illustrates crack propagation through incremental slicing, with each consecutive micrograph representing the 10th slice along the z-axis at 600 nm intervals. These images reveal cracks originating from the particle core and extending outward to the surface, emphasising the role of internal voids in initiating radial cracks and their significance in structural degradation during cycling. Parks et al.233 observed similar radial crack propagation patterns, highlighting the importance of internal voids in driving crack growth. 4.3 Results and Discussion 93 Figure 4.14: Cross-sectional SEM images of a polycrystalline (PC) NMC811 electrode after 300 cycles at 1C rate acquired using slice-and-view technique, with each image representing the 10th slice along the z-axis at 600 nm intervals. Scalebar represents 2 𝜇𝑚. As cracks extend from particle core to surface, they create pathways for electrolyte penetration, intensifying parasitic reactions and promoting irreversible lithium loss. This dual degradation mechanism accelerates capacity fade and structural failure, surpassing the impact of surface- localised cracking alone. Image processing reveals cracking severity is most pronounced at the particle core and decreases radially outward, consistent with established electrochemical cracking mechanisms. 4.3.2 Tortuosity Analysis: Dense vs Sparse Graphs Figure 4.15 compares tortuosity values derived from dense and sparse graphs for active material and pore phases in the calendering direction. Dense graphs generally yield higher tortuosity values for both phases due to their more complete representation of connectivity (as described in Section 4.2.4), capturing subtle structural features that sparse graphs may miss. 94 4.3 Results and Discussion Figure 4.15: Comparison of tortuosity values derived from dense and sparse graphs for (a) Active Material and (b) Pore in the throat-weighted analysis along direction of calendering. Error bars represent the standard deviation of the pathway tortuosity distribution reported by Dragonfly for the selected input-output boxes. Although precise computation times could not be determined due to concurrent workstation use, literature suggests analysing dense graphs can be up to thirty times more computationally intensive than sparse graphs.234 In resource-constrained scenarios, sparse graphs are particularly advantageous, mitigating memory limit risks while still delivering meaningful results. These findings align with Duan et al.,234 who reported that sparse graphs, when appropriately scaled, can reliably approximate tortuosity values while maintaining computational efficiency. Specifically, tortuosity values from dense graphs are approximately 1.1 times those from sparse graphs for active material and 1.2 times for pores. This highlights sparse graphs' practicality for resource-limited environments or large, memory-intensive datasets. Active Material Pore a. b. 4.3 Results and Discussion 95 4.3.3 Tortuosity Analysis: Throat weighted graphs The comparison between throat-weighted and non-throat-weighted graphs reveals significant differences in their ability to represent structural trends within porous media, as shown in Figure 4.16. Non-throat-weighted graphs, with equal edge weights, showed minimal variation across samples, masking calendering effects. In contrast, throat-weighted graphs amplified tortuosity in regions of reduced pore connectivity by incorporating throat dimensions. For uncalendered SC, throat-weighted tortuosity of active material was markedly higher, illustrating complex, constricted transport pathways. As calendering progressed, throat- weighted tortuosity values systematically decreased, a trend not captured in non-throat- weighted metrics, demonstrating that ignoring throat geometry can underestimate calendering's structural influence on transport properties. 235 These findings are consistent with prior research in pore-network modeling and image-based characterisation of porous materials. Studies show incorporating throat geometries into pore- network models leads to more accurate transport property predictions. Dong and Blunt236 demonstrated that accounting for throat dimensions and shape factors of micro-CT images improves correlation between simulated and experimental permeability measurements. Similarly, other researchers have emphasised that physically meaningful weighting schemes, which integrate geometric and topological characteristics of the pore network, yield more realistic and sensitive transport predictions.237,238 Figure 4.16: Comparison of tortuosity values derived from throat weighted and non-throat weighted for active material and pore along the direction of calendering. Error bars represent the standard deviation of the pathway tortuosity distribution reported by Dragonfly for the selected input-output boxes. a. b. Active Material Pore 96 4.3 Results and Discussion Commercial platforms like Dragonfly software enable both non-throat and throat-weighted calculations, allowing three-dimensional visualisation of constricted pathways within calendered electrodes. By accurately capturing regions of reduced connectivity, throat- weighted networks closely correlate with observed changes in transport parameters. Having established the value of throat-weighted analysis, the next section compares Dragonfly and TauFactor to assess how computational approach influences tortuosity accuracy. 4.3.4 Comparison of Taufactor and Dragonfly Calculating tortuosity using TauFactor for a dataset with 20 𝜇𝑚3 volume at 10 × 10 nm2 cross- section pixel size required multiple days. To address this computational demand, two strategies were employed: (1) calculating tortuosity in smaller regions using a sub-volume approach and (2) rebinning the dataset to coarser resolution. In the sub-volume approach, the dataset is divided into four smaller 5 × 5 𝜇𝑚2 regions (highlighted by red squares in Figure 4.8) with tortuosity calculated individually and then averaged. This significantly reduces computational time while retaining original voxel resolution. The full-volume method analyses a larger 10 × 10 𝜇𝑚2 domain as a single dataset (denoted by the blue dashed square in Figure 4.8), providing a broader view but requiring more computation. The sub-volume method works particularly well for SC electrodes, where 1 - 5 𝜇𝑚 diameter particles are well represented within 5 𝜇𝑚 sub-regions. Tortuosity values differ by only 1 - 7% from full-volume method, demonstrating reliability for SC electrodes. However, for PC electrodes with larger 5 - 10 𝜇𝑚 secondary particles, the sub-volume approach becomes less reliable, with discrepancies of 5 - 19% compared to full-volume calculations. The sub-volume method significantly reduces computation time: SC samples require approximately 65 minutes versus 184 minutes for full-volume calculation, while PC datasets take 149 minutes versus 209 minutes. Rebinning offers an alternative by downsampling voxel resolution. Two-fold rebinning increases voxel size from 10 nm × 10 nm to 20 nm × 20 nm, reducing total voxel count while retaining sufficient resolution for nanocracks and binder networks, as shown in Figure 4.17. Further increasing rebinning to three- or four-fold compromises accuracy by obscuring finer details, particularly in PC electrodes. For SC-25, two-fold rebinning resulted in only 4.8% tortuosity difference compared to the sub-volume method, while reducing computational time fivefold. Both the sub-volume approach and rebinning provide practical solutions for tortuosity calculations in high-resolution datasets. While the sub-volume method is particularly well- 4.3 Results and Discussion 97 suited for SC electrodes, it is less reliable for PC systems due to their larger particle sizes. Rebinning improves computational efficiency but must be carefully optimised to preserve critical structural features. Future work could explore increasing the number of sub-volumes beyond the current four or employing high-performance computing resources to enable full- volume analyses for larger datasets. Access to a supercomputer could facilitate routine full- volume calculations, ensuring robust and more comprehensive results reflective of the actual electrode structure. Figure 4.17: Comparative FIB-SEM images of single-crystal (top row) and polycrystalline (bottom row) electrodes at 25% porosity. Scale bar: 5 μm. Pore-phase tortuosity along the direction of calendering (through-plane) derived from TauFactor and Dragonfly for single-crystal (SC) and polycrystalline (PC) electrodes at porosities of 43%, 35% and 25%. (a) TauFactor tortuosity calculated using the full-volume approach (solid triangles) and the sub-volume approach (open triangles). (b) Comparison of Dragonfly dense-graph tortuosity (green circles) with TauFactor sub-volume tortuosity (open triangles). For the sub-volume TauFactor results, error bars represent the standard deviation across four 5 × 5 µ𝑚² subvolumes, each spanning the full reconstructed thickness; for the Dragonfly dense-graph results, error bars represent the standard deviation of the pathway tortuosity distribution reported by Dragonfly for the selected input-output boxes. TauFactor has been widely used over the past few years for calculating the tortuosity of lithium-ion battery (LIB) electrodes, establishing itself as a trusted method in the field. To assess the reliability of Dragonfly as a complementary tool, a comparison was conducted 5 𝜇𝑚 5 𝜇𝑚 1 2 3 4 98 4.3 Results and Discussion between results obtained using Dragonfly's sparse graph with throat-weighted method, and TauFactor's sub-volume approach. Both methods exhibit consistent trends in Figure 4.18, demonstrating their ability to capture the relationship between porosity and tortuosity. However, Dragonfly's dense, throat-weighted graphs yield consistently higher tortuosity values than those obtained using TauFactor, a factor of 1.7 - 2 for SCs and 1.1 - 1.5 for PCs. This discrepancy arises from differences in computational approaches.153,218 While efficient for large datasets, it averages transport properties across discrete regions, smoothing out bottlenecks in smaller sub-volumes. Dragonfly's throat-weighted method incorporates physical dimensions of narrow pathways, capturing localised transport barriers like constricted pores and microcracks. This sensitivity to geometric constraints leads to higher tortuosity values, particularly in polycrystalline electrodes, where structural heterogeneities are more prevalent. Figure 4.18: Comparison of tortuosity values calculated using the full-volume (solid triangles) and sub-volume (open triangles) approaches for single-crystal (SC) and polycrystalline (PC) samples at varying porosities (43%, 35%, and 25%) The error bar represents the standard deviation between the four sub-volumes. The comparison highlights the complementary strengths of both methods. TauFactor’s finite difference approach is robust and efficient, making it well-suited for porous materials, while Dragonfly excels in capturing geometric constraints and heterogeneities in more complex microstructures. Dragonfly’s user-friendly interface further enhances its accessibility, making it a practical alternative for researchers less familiar with coding. Together, these methods provide valuable insights into pore morphology and connectivity, reinforcing their utility for analysing structural changes in SC and PC electrodes under varying calendering conditions. a. b. Pore Pore 4.3 Results and Discussion 99 4.3.5 Symmetrical EIS Analysis and the Bruggeman Relationship The effects of decreasing porosity due to calendering on bulk electrode properties were investigated using EIS on symmetric cells under blocking conditions. Figure 4.19a presents EIS spectra for symmetric cells with varying porosity PC and SC NMC811, using 10 mM tetrabutylammonium perchlorate (TBAClO4) salt in 3:7 EC−EMC electrolyte. The ionic conductivity of the electrolyte is based on the reported literature221 value of 0.332 mS cm−1 used to calculate the tortuosity and MacMullin numbers of the electrodes. The EIS spectra show three key features: a high-frequency semicircle, a 45° line at medium frequencies, and constant-phase behaviour at low frequencies. Literature attributes the semicircle to electrode-current collector contact resistance, while the subsequent features represent impedance from ion and electron conduction through the porous cathode, described by the general transmission line model.221,239 Both PC and SC electrodes show consistent resistance trends across calendering conditions, with calendered electrodes exhibiting lower resistance than uncalendered counterparts. Compression forces particles and current collector into more intimate contact, lowering resistance until maximum connections form, at which point the semicircle may disappear. As semicircles remain visible across all samples, this limit hasn't been reached, suggesting potential for further calendering to improve particle contact. Later sections will explore why this is applicable to SC electrodes but not to PC electrodes. Figure 4.19: (a) Nyquist plot of symmetric cell assembled with polycrystalline (PC) and single crystal (SC) NMC811 under blocking conditions. (b) Calculated MacMullin numbers using the ionic resistance obtained from the Nyquist plot. As a rule, MacMullin numbers greater than 1 reflect lower ionic conductivity within the electrode compared to the free electrolyte.240 Figure 4.19b shows MacMullin numbers at varying porosity levels. For PC electrodes, the number increases from uncalendered to 25% a. b. 100 4.3 Results and Discussion porosity, confirming ionic conductivity decreases with reduced porosity due to restricted ion movement. This trend highlights that bulk mass transport becomes increasingly difficult at 25% porosity due to reduced pore volume, which restricts ion movement through the electrode. In contrast, SC samples exhibit a decrease in MacMullin number, from 8.87 in uncalendered samples to 6.84 at 35% porosity and 5.19 at 25% porosity. This behaviour deviates from the traditional Bruggeman relationship, which predicts increased transport resistance with reduced porosity. Several factors explain this: calendering likely promotes better particle alignment and packing of single crystals, creating more direct transport pathways.241 Compression may eliminate tortuous dead-end pores while preserving critical transport channels, as Thorat et al.35 in their study on LiFePO4. Additionally, single-crystal particles facilitate better stress distribution during calendering, mitigating bottleneck formation.242 Figure 4.20: The tortuosity-porosity relationship of (a) PC with Bruggeman fitting, (b) PC with Bruggemen fitting and pre-exponential factor and (c) SC. 4.3 Results and Discussion 101 Even PC samples show deviations from the theoretical Bruggeman relationship, as illustrated in Figure 4.20a. Using Dragonfly, TauFactor, and EIS, the best-fit Bruggeman exponents 𝛼 for PC samples are 1.96, 1.73, and 1.62, respectively, higher than the theoretical value of 1.5 for ideal spherical particles. This deviation arises from several factors. The Bruggeman model assumes perfectly spherical, non-overlapping particles, which oversimplifies the complex microstructure of real electrodes. Additionally, the presence of the carbon-binder domain introduces extra transport resistance, while variations in particle size and morphology create heterogeneous transport pathways. These assumptions are more applicable to high-porosity systems, whereas the calendering process introduces anisotropic pore networks that significantly affect transport properties.205 Literature suggests tortuosity-porosity relationships align well in 40-70% porosity range but diverge at lower porosities.211 Our study's range (25-43%) falls below this optimal range, contributing to observed deviations. At lower porosities, CBD influence becomes more pronounced. To better describe real electrode behaviour, modifications to the traditional Bruggeman relationship are needed, including a pre-exponential factor 𝛾 to account for morphology and composition variations and employing empirical Archie's law for improved fitting.35,212,243,244 Higher Bruggeman exponents exceeding 1.5 are often required to accurately capture the increased tortuosity observed in practical electrodes.35,245 While fitting with 𝛼 alone provides reasonable results for TauFactor and Dragonfly, it is less accurate for EIS, best fit curves illustrated in Figure 4.20b. When the pre-exponential factor is included, the α values align more closely with theoretical predictions. The 𝛾 and 𝛼 calculated is 1.09 and 1.89 for Dragonfly, 1.46 and 1.42 for Taufactor and 1.89, 1.09 for EIS, respectively. These findings align with Thorat et al.,35 who reported 𝛾 = 1.8 and 𝛼 = 1.53 for LiFePO4 and LiCoO2 electrodes. Chung et al.210 found NMC111 follows the Bruggeman relationship but consistently lies above it. Vijayaraghavan et al. 235 reported simulations of computer-generated electrodes matched the Bruggeman relationship, but when porosity clusters were introduced (as observed in FIB-SEM tomography Figure 2.9), tortuosity values matched experimental results. This indicates inhomogeneous particle packing increases tortuosity by 13–16% compared to uniformly packed electrodes, contributing to higher-than-expected Bruggeman estimations. These findings underscore the need for more sophisticated models that accurately account for the complexities of real electrode structures, including particle morphology, the carbon- binder domain (CBD), and anisotropic pore networks. Simple theoretical models, such as the traditional Bruggeman relationship, often fail to capture these intricacies and may significantly underestimate tortuosity in commercial electrodes. Refining tortuosity-porosity relationships is critical to achieving accurate predictions and optimised designs for lithium- ion batteries, particularly at industrially relevant lower porosities. 102 4.3 Results and Discussion The porosity range investigated in this study (25 - 43 %) falls below the optimal range for which the Bruggeman model is most applicable (40 - 70 %). This contributes to the observed deviations and highlights the limitations of relying solely on such models for electrode design. Relying solely on such models risks suboptimal electrode designs. Therefore, studying anisotropic transport behaviour is essential to understand and leverage directional transport properties, enabling more effective designs and improved battery performance. The tortuosity trends and Bruggeman relationship deviations discussed above are based on single EIS measurements for each electrode condition. While the observed trends are internally consistent and supported by complementary imaging-based measurements (Dragonfly and TauFactor), readers should note that the absolute tortuosity values and fitted parameters (Bruggeman exponent α, pre-exponential factor A) cannot be statistically validated without replicate measurements. The conclusions drawn from EIS data therefore represent preliminary observations that are strengthened by: 1. Consistency with microstructural observations from FIB-SEM, 2. Correlation with electrochemical performance trends (rate capability, impedance growth), 3. Qualitative agreement with independent tortuosity measurements from Dragonfly and TauFactor methods. Future work involving replicate symmetric cell measurements would enable rigorous statistical validation of the specific numerical values reported here and allow quantitative assessment of measurement uncertainty in Bruggeman parameters. 4.3.6 Electrochemical Implications The electrochemical performance of SC and PC electrodes under varying calendering conditions reveals distinct trends that underscore the structural advantages of SC electrodes. Detailed electrochemical analyses including rate capability tests, EIS measurements, GITT, HPPC, and long-term cycling data are presented and discussed in our associated paper. 114 For readers interested in understanding the comprehensive electrochemical characterisation and graphical representations of these results, please refer to that publication. Here, we provide a concise summary of the key findings as they relate to the microstructural observations in this study. Rate capability tests demonstrate that SC electrodes maintain stable C-rate performance across different packing densities, whereas PC electrodes suffer significant capacity loss at higher C- 4.3 Results and Discussion 103 rates, especially under heavy calendering. This highlights the superior ability of SC electrodes to retain transport pathways despite increased compression. EIS measurements at 50 % SOC further support these findings. SC electrodes show systematic reduction in charge transfer resistance (𝑅𝐶𝑇) and series resistance as porosity decreases. For PC electrodes, while 𝑅𝐶𝑇 values also decrease initially, heavily calendered PC-25 exhibits increased series resistance, attributed to calendering-induced cracking. These differences reflect SC electrodes' ability to withstand mechanical stress during calendering, maintaining more uniform and effective transport pathways. Table 4.1: Compilation of electrochemical test results carried out by Dr Kumar Raju for SCs and PCs at varying porosity levels. UC = Uncalendered 114 Single Crystal Polycrystalline UC 35% 25% UC 35% 25% 𝑹𝑰𝒐𝒏 / 𝛀 680.6 420.0 278.9 412.2 415.9 520.0 𝑹𝑪𝑻 / 𝛀 105.7 79.3 32.2 184.8 133.3 33.6 𝑹𝑯𝑭 / 𝛀 11.3 9.5 4.1 6.6 4.3 7.3 Capacity retention after 300 cycles / % 65 70 76 62 63 66 Capacity retention after 100 cycles / % 63 82 86 61 63 68 Hybrid Pulse Power Characterisation (HPPC) results reinforce the limitations of heavily calendered PC electrodes. PC-25 samples fail to achieve full depth of discharge after 200 cycles, coupled with a steady increase in resistance, signalling progressive degradation of transport pathways. Conversely, SC-25 samples exhibit lower polarisation growth and maintain better pulse performance, highlighting the robustness of SC electrodes under high compression. Finally, long-term cycling stability emphasises the superior durability of SC electrodes. After 300 cycles, SC-25 retains 76% of its initial capacity compared to 66% for PC-25. Notably, 104 4.3 Results and Discussion uncalendered SC electrodes perform worse than calendered SC electrodes, further demonstrating the positive impact of calendering on structural integrity and performance. In contrast, heavily calendered PC electrodes show a drastic 100% increase in 𝑅𝐶𝑇 within the first 100 cycles, reflecting the detrimental effects of increased surface area exposure to the electrolyte and associated cracking. These findings underscore the superior performance of SC electrodes, particularly under heavy calendering conditions, where their structural integrity and transport properties remain intact. Enhanced performance metrics, such as reduced impedance build-up, improved lithium-ion diffusivity, and stable rate capability, are consistently supported by EIS and HPPC data. In contrast, heavily calendered PC electrodes experience significant capacity fade due to increased surface area exposure to the electrolyte, which exacerbates cracking and accelerates impedance growth. 4.3.7 Tortuosity Anisotropy and Transport Mechanisms in SC and PC NMC811 Electrodes Analysis of tortuosity anisotropy in NMC811 electrodes reveals critical differences between SC and PC electrodes with significant performance implications. Figure 4.21 presents tortuosity measurements for the porous network using three complementary techniques: Dragonfly (green), TauFactor (orange), and symmetrical EIS (brown). Panels (a-c) show method-specific comparisons of through-plane (solid lines, dark colours) and in-plane (dashed lines, light colours) tortuosity, while panels (d-e) present direction-specific comparisons facilitating direct comparison across techniques. The vertical dotted line separates SC and PC sample types. EIS measurements, performed by Dr Kumar Raju, were only conducted in the through-plane direction due to the symmetric cell configuration. Understanding the error bars in Figure 4.21 and Figure 4.22 is essential for proper interpretation. For TauFactor (orange), error bars represent the standard deviation from four independent 5 × 5 × 15 𝜇𝑚³ subvolumes, quantifying spatial heterogeneity of transport properties. For Dragonfly (green), error bars represent the standard deviation of tortuosity across all calculated pathways, capturing variability in pathway complexity, where larger values indicate coexistence of highly tortuous and more direct routes. For EIS (brown), no error bars are shown because only one symmetric cell was measured per condition; however, the trends are broadly consistent with both imaging-based methods. 4.3 Results and Discussion 105 Figure 4.21: Tortuosity measurements for porous network across different sample types and measurement methods. (a-c) Method-specific comparisons showing through-plane (solid lines, dark colours) and in-plane (dashed lines, light colours) tortuosity for (a) Dragonfly, (b) Taufactor, and (c) EIS measurements across SC (uncompressed, 35%, 25% compressed) and PC (uncompressed, 35%, 25% compressed) samples. (d-e) Direction-specific comparisons showing (d) through-plane and (e) in-plane tortuosity for all three methods (Dragonfly: green circles, Taufactor: orange triangles, EIS: brown squares). Error bars represent standard deviation. The vertical dotted line separates SC and PC sample types. Note: EIS measurements were only performed in the through-plane direction. Here, tortuosity anisotropy is defined as the ratio of through-plane tortuosity over in-plane tortuosity. For anisotropy ratios (Figure 4.22), error propagation followed: 𝜎𝑎𝑛𝑖𝑠𝑜𝑡𝑟𝑜𝑝𝑦 = 𝜏𝑡ℎ𝑟𝑜𝑢𝑔ℎ−𝑝𝑙𝑎𝑛𝑒 𝜏𝑖𝑛−𝑝𝑙𝑎𝑛𝑒 × √( 𝜎𝑡ℎ𝑟𝑜𝑢𝑔ℎ−𝑝𝑙𝑎𝑛𝑒 𝜏𝑡ℎ𝑟𝑜𝑢𝑔ℎ−𝑝𝑙𝑎𝑛𝑒 ) 2 + ( 𝜎𝑖𝑛−𝑝𝑙𝑎𝑛𝑒 𝜏𝑖𝑛−𝑝𝑙𝑎𝑛𝑒 ) 2 Equation 19 where 𝜎𝑡ℎ𝑟𝑜𝑢𝑔ℎ−𝑝𝑙𝑎𝑛𝑒 and 𝜎𝑖𝑛−𝑝𝑙𝑎𝑛𝑒 are the standard deviations of through-plane and in-plane tortuosity measurements, respectively. a. b. c. d. e. 106 4.3 Results and Discussion The relatively large error bars for some samples, particularly PC-25, reflect both inherent microstructural heterogeneity and challenges in quantifying transport in heavily calendered or cracked structures. Examining the detailed anisotropy evolution for each electrode type reveals how calendering creates different transport architectures in SC versus PC electrodes. Beginning with the pore phase in SC electrodes, tortuosity values (Figure 4.21a,d) remain relatively constant in both directions for SC-UC and SC-35 according to both Dragonfly and TauFactor. The anisotropy ratios (Figure 4.22a) exceed unity for these conditions, indicating slightly higher through- plane tortuosity as a consequence of gravity-driven particle settling during fabrication. However, calendering to SC-25 fundamentally changes this behaviour: through-plane tortuosity decreases while in-plane tortuosity increases, creating an anisotropy ratio of approximately 0.75. This transition to preferential through-plane transport facilitates enhanced ionic transport from separator to active material.205,241 EIS measurements strongly support this trend, showing pronounced through-plane tortuosity reduction from 7.63 (SC- UC) to 5.12 (SC-35) to 2.60 (SC-25). The convergence across three independent techniques, despite different length scales (nanoscale FIB-SEM vs. macroscale EIS) and physical principles (geometric path analysis vs. electrochemical impedance), supports the findings that SC electrodes develop preferentially aligned pore networks under heavy calendering.246,247 Turning to the SC’s active material phase, SC electrodes show initial discrepancies between Dragonfly and TauFactor for uncalendered samples (Figure 4.22b), likely reflecting challenges in skeletonising the poorly connected electronic network. However, both methods demonstrate reduced anisotropy with calendering, converging at SC-35 to indicate nearly isotropic transport (anisotropy ≈ 1). Further calendering to SC-25 promotes preferential through-plane electron transport (anisotropy < 1), creating more direct pathways to the current collector. This enhanced alignment, combined with substantial charge transfer resistance reduction from 105.7 𝛺 (SC-UC) to 57.8 𝛺 (SC-35) to 32.2 𝛺 (SC-25), demonstrates concurrent optimisation of ionic and electronic pathways. GITT measurements showing improved lithium diffusivity and HPPC tests demonstrating lower polarisation growth for SC-25 are consistent with these optimised transport properties. When consolidating findings from both phases, SC electrodes exhibit consistent downward tortuosity trends with increasing calendering across all measurement techniques. Although intragranular cracks are observed in calendered SC electrodes (Figure 4.11 - Figure 4.13), they do not significantly impair the overall transport network, as evidenced by superior rate capability, lower polarisation resistance, and stable cycling. SC-25 achieves preferential transport in both phases, with anisotropy ratios below unity enhancing both ionic and electronic efficiency in the calendering direction. 4.3 Results and Discussion 107 Figure 4.22: Tortuosity anisotropy in NMC811 electrodes for pore and active-material phases. Tortuosity anisotropy is defined as 𝜏through-plane/𝜏in-plane . (a) Pore-phase anisotropy and (b) active-material-phase anisotropy, extracted using Dragonfly graph-based analysis and TauFactor for each sample (SC-UC, SC-35, SC-25, PC-UC, PC-35, PC-25). The dashed horizontal line at unity indicates isotropic transport (𝜏through = 𝜏in), while values above (below) 1 indicate higher tortuosity in the through-plane (in-plane) direction. The dotted vertical line separates single-crystal (SC) and polycrystalline (PC) electrodes. In contrast to the beneficial anisotropy evolution in SC electrodes, PC electrodes exhibit fundamentally different behaviour. Within the pore phase in PC, tortuosity is consistently higher in-plane than through-plane across all calendering conditions (Figure 4.21a,d), yielding anisotropy values below unity (Figure 4.22a). This persistent anisotropy suggests a. b . 108 4.3 Results and Discussion particle orientation is established primarily during slurry coating and drying rather than calendering, as the larger, more spherical PC secondary particles have limited ability to reorient under pressure. This aligns with Usseglio-Viretta et al.205 for NMC532 and Ebner et al.,83 who noted that anisotropy relates to fractured fragment orientation rather than whole particle alignment. A notable tortuosity spike occurs from 35 % to 25 % porosity, evident in both imaging methods, reflecting severe pore connectivity reduction from compaction and extensive intergranular cracking (Figure 4.11-Figure 4.13). The large PC-25 error bars indicate substantial spatial heterogeneity, with some regions showing intact networks while others are severely disrupted. Interestingly, EIS shows relatively constant pore tortuosity (2.0 - 2.2) across all PC conditions. This apparent discrepancy reflects different length scales, where EIS provides bulk-averaged measurements over approximately 1 𝑐𝑚², while FIB-SEM samples specific 10 to 20 𝜇𝑚 regions. The macroscopic measurement averages over both cracked and intact regions, smoothing local heterogeneities. This highlights the complementary nature of the techniques, where imaging reveals local degradation while EIS confirms bulk transport is not catastrophically impaired because alternative pathways remain available. Within the active material phase of PC electrodes, they exhibit conflicting anisotropy trends between methods (Figure 4.22b). While uncalendered samples show reasonable agreement (Dragonfly ≈ 1.5, TauFactor ≈ 1.2), calendered samples yield opposing results: Dragonfly suggests enhanced through-plane transport (anisotropy < 1), whereas TauFactor indicates preferential in-plane transport (anisotropy > 1). These inconsistencies, combined with large error bars spanning 0.5–2.7 for Dragonfly, indicate fundamental challenges in characterising heavily fractured structures. Contributing factors include discontinuous solid-phase regions that are difficult to skeletonise, heterogeneous carbon-binder distribution creating ambiguous connectivity, and computational differences between graph-based and finite-difference approaches when analysing partially fractured networks. The large uncertainties reflect physical reality rather than measurement failure, as heavy calendering creates severe heterogeneity where some regions remain intact while others are extensively fractured. Future work using plasma FIB technology, enabling fields of view exceeding 100 𝜇𝑚² while maintaining nanoscale resolution, would provide more statistically representative sampling. Overall, PC electrodes show complex relationships between calendering and transport. Active material tortuosity improves with calendering, and charge transfer resistance decreases substantially from 184.8 Ω (PC-UC) to 133.3 Ω (PC-35) to 33.6 Ω (PC-25). However, these electronic improvements are offset by pore phase deterioration, with ionic resistance rising significantly from 415.9 Ω (PC-35) to 520.0 Ω (PC-25), consistent with increased ionic transport limitation. GITT measurements show reduced lithium diffusivity for PC-25, while HPPC tests reveal PC-25 fails to achieve full depth of discharge due to excessive polarisation. While moderate calendering to PC-35 provides beneficial compaction with manageable damage, 4.3 Results and Discussion 109 heavy calendering to PC-25 causes intergranular fracture that creates ionic bottlenecks overwhelming any electronic improvements. The tortuosity trends from all three methods correlate well despite different absolute values. Imaging-based techniques provide directional tortuosity and anisotropy analysis, while EIS provides macroscopic validation under realistic conditions. For SC electrodes, the pronounced EIS trends supports that microscale improvements translate to macroscopic performance. For PC electrodes, the relatively constant EIS values despite severe local cracking reveal a critical multi-scale insight: alternative transport pathways at larger length scales partially compensate for local structural degradation. Table 4.2 provides a comprehensive overview of these techniques with their advantages and limitations. 110 4.3 Results and Discussion Table 4.2: Comparison of tortuosity calculation methods for NMC811 electrodes Taufactor Dragonfly EIS Algorithmic approach Uses finite-difference method to calculate tortuosity from voxel-based segmented images Processes 3D tomographic data through advanced image segmentation and skeletonisation Electrochemical impedance measurements to estimate tortuosity using equivalent circuits Advantages • Can analyse multiple directions • Easy to use for basic tortuosity calculations • Fast computation • Free, open-sourced • Widely used in literature for comparison • Comprehensive toolkit with advanced 3D segmentation and machine learning integration • Can process large datasets with high accuracy • Free software for academic usage • Allows integration with third-party scripts and workflows • Non-destructive • Provides broader electrochemical insights (eg: resistances and capacitances) • Widely used in literature for comparison • Operates in real-time for in situ or operando studies Disadvantages • Require 3D imaging data • Limited by image resolution, and tomography volume • Requires coding • Computationally intensive • Time intensive to acquire images • Costly software • Limited by image resolution, and tomography volume • Limited to through-plane measurements • Interpretation of results is highly model-dependent 4.3 Results and Discussion 111 Based on the tortuosity results, Figure 4.23 presents a conceptual schematic of how microstructural evolution and transport mechanisms for ions and electrons may differ between SC and PC electrodes under varying calendering conditions. In PC electrodes, the collapse of large voids at low calendering (35% porosity) initially aids Li+ ion movement by reducing tortuous pathways. However, at higher calendering (25% porosity), the formation of intergranular cracks disrupts connectivity, hindering ionic transport. In contrast, the electronic transport benefits from better particle-to-particle contact, facilitated by the redistribution and compression of the carbon black within the carbon-binder domain (CBD), which enhances electronic conductivity as calendering increases. Conversely, SC electrodes exhibit enhanced structural alignment and reduced tortuosity for both active material and pore phases as calendering increases, supporting more efficient ionic and electronic transport. The uniform compression and alignment of SC particles promote improved pathways for both Li+ ions and electrons. The presence of well-distributed carbon black within the CBD further supports enhanced electron transport in SC electrodes, ensuring robust connectivity and reduced resistance. This highlights the critical role of CBD redistribution during calendering in optimising both ionic and electronic transport properties, particularly in SC electrodes. In contrast, PC electrodes, made of larger, spherical particles, exhibit heterogeneous deformation under calendering, leading to inter-particle displacement, uneven stress distribution, and greater pore connectivity loss, especially at higher calendering intensities. Their spherical geometry limits structural alignment and results in minimal tortuosity anisotropy, with similar values across all three directions. Furthermore, the larger size of PC particles makes them more susceptible to cracking and structural degradation under compression. By comparison, the better packing and alignment of smaller SC particles enable enhanced transport properties and improved mechanical stability. 112 4.3 Results and Discussion Figure 4.23: Schematic illustration of ion and electron transport pathways in single-crystal (SC) and polycrystalline (PC) NMC811 electrodes at different porosities. Top row: PC electrodes showing increased intergranular cracking and tortuous pathways with calendering. Bottom row: SC electrodes demonstrating enhanced alignment and reduced tortuosity with calendering, promoting more direct ionic and electronic transport pathways. Pink arrows indicate ion (Li+) movement, and yellow arrows show electron transport through the carbon black network (blue) through grain boundaries or from one particle to another The anisotropic behaviour observed in SC electrodes has significant implications for battery performance. At 25% porosity, preferential pore alignment in the through-plane direction facilitates faster ion transport across the electrode thickness, enhancing ionic movement from the separator to the active material. This reduces concentration gradients during cycling and promotes uniform utilisation of the active material. Calendering also induces anisotropy in the active material, reflecting microstructural reorganisation that improves particle-to-particle contact, strengthens electrical connectivity, and enhances mechanical stability, thereby lowering the risk of particle fracture. These aligned pathways reduce resistance, mitigate concentration polarisation, and improve electrolyte penetration, findings supported by experimental performance tests. Furthermore, the smaller particle size in SC electrodes shortens diffusion paths and improves reaction kinetics, enabling superior performance at high C-rates. In contrast, PC electrodes, with larger particles and minimal structural alignment, suffer from longer diffusion lengths and increased polarisation, particularly under high calendering pressures. Optimising the balance between porosity and thickness in SC electrodes offers a promising route to simultaneously enhance transport properties and mechanical robustness, ultimately improving rate capability and overall battery performance. 4.4 Conclusion 113 4.3.8 Electrode Optimisation Optimisation strategies should be tailored to particle morphology and structural characteristics. For SC electrodes, calendering to 25% porosity provides optimal performance by maintaining beneficial alignment of both phases. At this compression level, preferential transport pathways enhance both ionic and electronic conductivity, while intragranular cracks does not significantly impair performance. SC particles' natural alignment tendency during manufacturing can be leveraged through gravity-assisted alignment during slurry coating. Since 25% porosity outperforms 35% for SC electrodes, further calendering below 25% could potentially yield additional improvements. However, this must balance against particle fracture risk and restricted ionic transport. Future studies should explore 20-25% porosities to determine if optimal balance exists at even higher compression levels. PC electrodes require a more conservative approach with moderate calendering to 35% porosity yielding optimal results. At this level, electrodes benefit from improved particle contact while preserving critical pore networks, maintaining stable ionic resistance around 415.9 Ω. It is crucial to avoid calendering beyond this point, as excessive compression leads to severe intergranular cracking and deteriorating transport properties. During optimisation, several key metrics should be monitored: charge transfer resistance indicates electronic pathway optimisation, MacMullin number provides ionic transport efficiency insight, and anisotropy ratios help assess transport pathway development, particularly important for SC electrodes. The ultimate goal is balancing porosity reduction against maintaining efficient transport properties, recognising each electrode type requires a distinct approach to achieve optimal performance while preserving structural integrity. This careful consideration of particle characteristics in calendering protocols is essential for developing high-performance lithium-ion battery electrodes. 4.4 Conclusion This chapter developed and validated a multi-method framework for quantifying microstructure-transport relationships in calendered NMC811 electrodes, combining FIB- SEM tomography with complementary tortuosity analysis approaches. The workflow integrates high-resolution 3D acquisition, machine learning-based segmentation using a 3D U-Net, and direction-resolved tortuosity quantification, addressing key barriers that have historically limited the routine use of FIB-SEM tomography for battery electrodes. The supervised segmentation differentiates active material, CBD and pore space while suppressing shine-through artefacts. Critically, this work quantifies tortuosity for both the pore network (ionic transport) and the percolating solid network (solid-phase/electronic transport) in parallel, whereas many studies focus primarily on pore-phase transport. 114 4.4 Conclusion Considering both networks independently indicates that optimising electrode performance requires concurrent improvement of ionic and electronic transport pathways. A critical comparison of three distinct tortuosity approaches clarified their relative strengths and limitations. Dragonfly’s throat-weighted graph analysis enables computationally efficient tortuosity estimation with results comparable to dense-graph calculations, but can be challenged by severely fractured structures where connectivity becomes ambiguous. TauFactor solves the diffusion equation directly on segmented microstructures without skeletonisation and is therefore robust for complex networks; subvolume sampling is sufficient for SC electrodes but introduces larger variability for heterogeneous PC electrodes. Symmetric-cell EIS measurements (performed by Dr Kumar Raju) provide macroscopic context by probing transport over the full electrode thickness and a large electrode area, indicating that electrode-scale transport can remain functional even when local fracture is observed in the tomographic volumes. Overall, the three methods yield convergent trends for the more structurally homogeneous SC electrodes, while larger divergences emerge for heavily damaged PC-25 electrodes. The central finding is that particle morphology governs calendering response. SC electrodes exhibit progressive microstructural alignment with calendering, achieving anisotropy ratios below unity at 25% porosity that favour through-plane transport in both pore and solid networks. Consistent reductions in resistive metrics (such as charge-transfer resistance and MacMullin number) support balanced pathway optimisation, and intragranular cracking does not appear to strongly impair bulk transport. In contrast, PC electrodes exhibit competing effects where calendering improves solid-phase contact but, at 25% porosity, extensive intergranular cracking and pore-network disruption increase ionic transport limitation. These observations motivate morphology-dependent processing windows: SC electrodes benefit from heavier calendering to ~ 25 % porosity (or lower), whereas PC electrodes require more conservative calendering to ~35 % porosity to avoid pore-network degradation. The measurements presented here are based on single samples per condition. While convergence across independent techniques supports the qualitative trends, rigorous statistical validation requires replicate sampling. Future work should prioritise multiple FIB- SEM acquisitions from different regions to quantify inter-sample variability, replicate symmetric-cell measurements (at least 3 cells per condition) to enable uncertainty estimates and statistical comparison, and larger tomographic volumes enabled by plasma FIB systems to provide more representative sampling of heterogeneous PC electrodes. Together, these steps would strengthen the statistical basis of the framework and support translation into design guidance for electrode manufacturing. Chapter 5 115 Chapter 5 Probing Charging Protocol-Induced Redox Transformations in NMC811 Using EELS This chapter develops and validates a multi-Gaussian fitting methodology for extracting quantitative oxidation state information from EELS spectra acquired on non-monochromated microscopes. Where instrumental broadening (~1.5 eV) typically obscures fine spectral features used for redox analysis, the fitting approach enables reliable determination of L3/L2 white-line ratios and oxygen K-edge ΔE parameters from overlapping core-loss edges. The methodology proved robust across datasets with energy calibration shifts of up to 4 eV, with processing times under five minutes per spectrum image. The analysis code has been made publicly available on GitHub to support reproducibility and broader adoption. The methodology is applied to investigate how charging protocol affects degradation in NMC811 cathodes. Three protocols were compared: a standard constant-current constant-voltage (CCCV) baseline, LowStart (which ramps current upward, extending the high-voltage constant-voltage phase), and LowEnd (which tapers current downward, reducing high-voltage exposure). Electrochemical characterisation by collaborators showed that LowEnd maintains superior capacity retention (~93% after 600 cycles) compared to CCCV (~85%) and LowStart (~78%). Spatially-resolved EELS mapping reveals the chemical origin of these performance differences. LowStart samples exhibit a ~ 20 nm surface reconstruction layer characterised by Ni3+ to Ni2+ reduction and diminished metal-oxygen hybridisation, while LowEnd samples show a thinner ~10 nm degradation zone with better-preserved bulk redox states. These spectroscopic findings were corroborated by atomic- resolution HAADF-STEM imaging, which directly visualised the rock-salt – spinel - layered phase progression from particle surfaces into the bulk. Notably, LowEnd samples exhibited more extensive intergranular cracking than LowStart samples, yet maintained superior electrochemical performance. This partial decoupling of mechanical and chemical degradation demonstrates that surface reconstruction, driven by extended high-voltage exposure, dominates capacity loss over mechanical damage. 116 5.1 Introduction to Charging Protocols and EELS Analysis for NMC811 Degradation Acknowledgement: The charging protocol design, electrochemical impedance spectroscopy (EIS), and long-term cycling experiments presented in this chapter were carried out by Dr Alexander Dimitrijevic at University College London. The reference EELS spectra for Mn²⁺, Mn3+, and Mn4+ were acquired by Dr Demie Kepaptsoglou at SuperSTEM. Mr Kelvin Chan assisted with the implementation and troubleshooting of the Gaussian fitting code. Their contributions were instrumental in enabling the comparative redox analysis presented in this study. The author developed the multi-Gaussian fitting methodology for EELS quantification described in this chapter and performed all TEM sample preparation, electron microscopy imaging, and EELS data acquisition and analysis presented herein. 5.1 Introduction to Charging Protocols and EELS Analysis for NMC811 Degradation The high nickel content in NMC811 makes them more chemically reactive and structurally unstable, particularly under high-voltage cycling conditions16. One of the main degradation pathways involves surface reconstruction of the layered structure into spinel and rock-salt phases, which significantly impedes lithium-ion transport59. These degradation processes span multiple length scales, from atomic-level changes in oxidation state to macroscopic declines in cell performance. Traditional performance metrics such as capacity retention and impedance growth are useful indicators of cell health, but they provide limited insight into the specific degradation mechanisms occurring within the electrode. In contrast, advanced characterisation techniques like electron energy loss spectroscopy (EELS) offer the ability to probe chemical changes at the atomic scale. However, these nanoscale observations must be carefully contextualised within broader electrochemical behaviour to draw conclusions relevant to practical battery use. Among the various factors influencing battery degradation, the charging protocol has emerged as a particularly accessible lever for extending cell lifetime. Unlike modifications to material composition or electrolyte formulation, adjustments to the current profile can be implemented without changes to cell construction. Recent studies have demonstrated that tailoring the current throughout the state-of-charge (SoC) range can meaningfully influence ageing outcomes, motivating detailed investigation of protocol-specific degradation pathways.248–250 To support this investigation, it is helpful to clarify two important concepts: C-rate and state of charge (SoC). The C-rate describes how quickly a battery is charged relative to its capacity. A 1C rate means the battery is fully charged in one hour, for example, a 5 Ah cell charged at 5 A. A 0.5C rate (2.5 A in this case) would take two hours. This normalisation makes it easier 5.2 Electrochemical Analysis 117 to compare performance across batteries of different sizes. SoC, on the other hand, refers to how full the battery is, expressed as a percentage from 0% (fully discharged) to 100% (fully charged). Different SoC regions stress the battery differently, with degradation typically more severe at the low and high ends of the charge window. This chapter investigates how three different charging protocols, constant current-constant voltage (CCCV), LowStart, and LowEnd, influence degradation in commercial 21700-format lithium-ion cells containing NMC811 cathodes and SiOₓ-graphite anodes. CCCV serves as the baseline protocol, applying a fixed current followed by a voltage hold. The LowStart protocol begins with a lower current and ramps up toward the end of charging, while the LowEnd protocol does the reverse, starting with a high current and tapering off before entering the voltage hold phase. All protocols were carefully designed to deliver the same total charge over the same duration, with an average C-rate of 0.5C, ensuring a fair comparison. Through a combination of electrochemical testing, impedance spectroscopy, and spatially resolved EELS, this study examines how the current profile affects surface reconstruction, redox evolution, and mechanical integrity in the cathode. This integrated, multiscale approach enables direct correlation between nanoscale chemical changes and bulk cell performance. Specifically, this chapter aims to: (i) evaluate how different charging profiles influence transition metal reduction and oxygen bonding at the nanoscale, (ii) correlate these chemical changes with electrochemical outcomes, and (iii) introduce a generalised, open-source Gaussian fitting tool to support the broader EELS community. Together, these contributions offer a deeper understanding of how tailored charging strategies can mitigate degradation and extend battery lifetime, without modifying cell materials. 5.2 Electrochemical Analysis The electrochemical characterisation and analysis presented in this section, including long- term cycling data, electrochemical impedance spectroscopy, and the design of the charging protocols, was performed by Dr Alexander Dimitrijevic at University College London. The author acknowledges these contributions and directs readers to Dr Dimitrijevic's thesis251 for comprehensive electrochemical analysis and experimental details. The following section provides a brief summary of the electrochemical results to contextualise the EELS analysis presented in Section 5.3. 118 5.2 Electrochemical Analysis 5.2.1 Charging Protocols The three charging protocols used in this study are illustrated schematically in Figure 5.1. In all cases the cells were charged to the same upper cut-off voltage and included a constant-voltage (CV) hold; the protocols differ in how the constant-current (CC) portion of the charge is distributed across state-of-charge windows. This design allows the effect of protocol-dependent high-SoC dwell to be probed while keeping the overall charge throughput comparable. The LowEnd protocol front-loads the CC charge by applying a higher current through the majority of the SoC window before tapering to a lower current as the cell approaches the upper cut-off voltage. By reducing the current in the final stage of charge, the overpotential and time spent close to the maximum voltage are decreased, which is expected to suppress high-voltage parasitic reactions and mitigate surface reconstruction. Figure 5.1: Schematic representation of all variable current protocols used, the C-rate applied per interval and how much time it would take to complete 0.5C equivalent. (a) is the commonly used CCCV, (b) the LowEnd, (c) the LowStart. This is only representative for the charge sequence. Protocols set by Dr Alexander Dimitrijevic. 5.2 Electrochemical Analysis 119 In contrast, the LowStart protocol redistributes the CC charge towards later stages of charge by applying a lower current at low SoC and a higher current at intermediate SoC. This current history increases the likelihood of an extended CV exposure at high voltage, deliberately accentuating high-SoC dwell conditions that accelerate surface chemical changes in Ni-rich layered oxides. While minor approximations are inevitable when discretising a continuous charge into SoC windows, the protocols were implemented so that the total programmed CC charge remained comparable. This ensures that the primary variable between protocols is the distribution of current near the end of charge and the associated duration of the CV hold, enabling a direct comparison of the resulting chemical and structural degradation. 5.2.2 Electrochemical Impedance Spectroscopy (EIS) Electrochemical impedance spectroscopy (EIS) was measured at selected states of charge to track protocol-dependent changes in resistive and transport processes and to provide an electrochemical context for the spectroscopic observations. In this chapter, only the key trends in (i) the high-frequency resistance, (ii) the mid-frequency charge-transfer response and (iii) the low-frequency transport contribution are summarised. Full experimental details (instrumentation, SoC points, circuit models and fitting procedures) are reported in Dr Dimitrijevic’s electrochemistry thesis,251 which should be consulted for quantitative parameter values and additional spectra. 5.2.3 Electrochemical Results Long-term cycling showed clear protocol dependence in capacity retention over 600 cycles. LowStart exhibited the greatest fade (≈78% capacity retention), the standard CCCV protocol showed intermediate behaviour, and LowEnd retained the highest capacity (≈93%). The full cycling curves and replicate statistics are reported in Dr Dimitrijevic’s electrochemistry thesis; Table 5.1 provides a compact summary for context. The relative ranking indicates that reducing high-voltage exposure improves cycle life: LowEnd, which limits high-SoC dwell, retained the most capacity, whereas LowStart, which increases high-voltage dwell, degraded fastest. These electrochemical trends are used here to frame the subsequent EELS analysis rather than to provide a full electrochemical characterisation. 120 5.3 Quantitative EELS Mapping of Redox States in NMC811 Table 5.1: Start and end discharge capacity (normalised) for graphite/NMC811 full cells cycled at ambient temperature under CCCV, LowStart and LowEnd protocols. Electrochemical data provided by Dr Alexander Dimitrijevic; full electrochemical characterisation available in Dr Dimitrijevic's thesis.251 Protocol Initial Capacity (Normalised) Final Capacity (Normalised) CCCV 1 0.85 LowStart 1 0.78 LowEnd 1 0.93 Electrochemical impedance spectroscopy revealed protocol-dependent resistance evolution. The LowStart protocol exhibited the most significant impedance growth across all states of charge, with ohmic and charge transfer resistances approximately doubling after 545 cycles. In contrast, LowEnd cells maintained substantially lower resistances, consistent with reduced interfacial degradation attributed to shortened CV phase duration. These electrochemical trends, detailed in Dr Dimitrijevic's thesis,251 provide the macroscopic context for the nanoscale EELS analysis that follows. Overall, the electrochemical datasets indicate that current distribution near the end of charge strongly influences both impedance growth and long-term capacity retention. In the remainder of this chapter, these macroscopic trends are interpreted using spatially resolved EELS to identify the associated changes in transition-metal redox behaviour and oxygen hybridisation at the particle surface. 5.3 Quantitative EELS Mapping of Redox States in NMC811 5.3.1 Redox Chemistry in Discharged NMC811 Cathode Material In the fully discharged (lithiated) state, the electronic structure of NMC811 is governed by the characteristic oxidation states of its transition metals26. Nickel is the primary redox-active species, typically distributed between Ni2+ and Ni3+. Manganese remains largely in the Mn4+ state and serves as a structural stabiliser, while cobalt exists predominantly as Co3+, with limited redox involvement within the operating voltage window.252 5.3 Quantitative EELS Mapping of Redox States in NMC811 121 Figure 5.2: EELS spectrum of NMC811 showing the characteristic absorption edges of oxygen (O-K1, O-K2), manganese (Mn-L2, Mn-L3), cobalt (Co-L2, Co-L3,), and nickel (Ni-L2, Ni-L3), used to analyse oxidation states Oxygen anions (O2-) contribute via hybridisation with the transition metal 3d orbitals, forming covalent bonds that are reflected in the fine structure of the oxygen K-edge. In the discharged state, this hybridisation is reduced, leading to a lower pre-edge intensity in the O K-edge due to the filled 3d states.252,253 The characteristic EELS absorption edges for Ni, Mn, Co, and O are shown in Figure 5.2. 5.3.2 Experimental Limitations in Oxidation State Discrimination A key challenge in EELS-based analysis of NMC811 is the limited energy resolution of non- monochromated systems. In this study, the Spectra-300 microscope exhibited a full width at half maximum (FWHM) of approximately 1.6 eV for the zero-loss peak under the parameters detailed in Section 3.4.2. In transition metal L-edges, multiplet splitting typically manifests as distinct peaks within the L3 edge that correspond to different electronic transitions. This splitting originates from electron-electron interactions within the partially filled 3d orbitals and reflects the complex electronic configuration of transition metals. Specifically, the multiplet splitting in Ni L3 represents the energy difference between different final states that arise from various coupling arrangements of the 2p core hole (created during the EELS excitation) with the 3d valence electrons. The exact pattern and energy separation of these multiplets is highly sensitive to the oxidation state, spin state, and local coordination environment of the nickel atoms, making it a valuable fingerprint for oxidation state analysis when properly resolved.126,252 Mn-L3 Mn-L2 Co-L3 Co-L2 Ni-L3 Ni-L2 O-K1 Pre-peak O-K2 Main peak 122 5.3 Quantitative EELS Mapping of Redox States in NMC811 As illustrated in Figure 5.3a, monochromated EELS (FWHM ~0.25 eV) clearly resolves this splitting in the Ni-L3 edge, revealing two peaks, Ni-L3,high and Ni-L3,low, separated by 1.7 eV. In contrast, our non-monochromated data (Figure 5.3b) captures only a single broadened peak with a slight shoulder, precluding direct visual discrimination of distinct electronic states. While finer dispersion settings (e.g. 0.05 eV/channel) can potentially reveal such features even with non-monochromated instruments, the 0.3 eV/channel dispersion used in this study was insufficient. Figure 5.3: (a) Monochromated EELS spectrum of Ni2+ showing clear multiplet splitting in the Ni-L3 edge, with distinct peaks corresponding to Ni-L3,high and Ni-L3,low, separated by approximately 1.7 eV. The high energy resolution (FWHM ~0.25 eV) enables the resolution of fine electronic structure features. Data provided by colleague, Mr Wei Huang (b) Non- monochromated EELS spectrum of Ni, acquired with 0.3 eV/channel dispersion, where the Ni-L3 edge appears as a single broadened peak with no resolvable multiplet splitting High-resolution EELS with reduced dispersion and monochromation could, in theory, resolve subtle spectral features such as multiplet splitting in the Ni-L3 edge254,255. However, these improvements come with significant trade-offs. Reducing the energy dispersion from the 0.3 eV/channel256 used in this study to finer settings (e.g. 0.05 eV/channel) increases acquisition time by approximately sixfold and reduces electron counts per channel, thereby compromising the signal-to-noise ratio (SNR). To compensate the low SNR, one might increase the beam current or exposure time, but this substantially elevates the risk of beam- induced damage, including transition metal reduction and oxygen loss. Additionally, finer dispersion narrows the total spectral window (typically to ~100 eV), making it impractical to capture both oxygen and transition metal edges in a single scan. Multiple sequential acquisitions would be needed to cover the full spectral range from the O K-edge to the Ni L- edge, further increasing the cumulative electron dose and the likelihood of artefactual changes during data collection. 5.3 Quantitative EELS Mapping of Redox States in NMC811 123 While monochromation can significantly enhance energy resolution, with a record energy resolution of 4.2 meV,257 it also substantially decreases beam current, often by one to two orders of magnitude. This would further degrade the SNR unless acquisition times are extended which could in return lead to beam-induced damage.255 Moreover, monochromated systems are less accessible due to their increased operational complexity, high cost, and limited availability, as they are not present in all research facilities.26,27,126,127,129,258,259 These constraints make them less suitable for routine use on beam-sensitive materials such as NMC811. Given these practical limitations, this study employed a non-monochromated system with a moderate dispersion of 0.3 eV/channel to balance spectral resolution with dose efficiency. Although this configuration does not resolve multiplet splitting, it allows simultaneous acquisition of oxygen and transition metal edges within a single scan and reasonable exposure time, minimising beam-induced damage. To address the resulting spectral broadening, a custom multi-Gaussian fitting routine was developed to efficiently deconvolute overlapping L3 and L2 features. This tailored approach enables accurate extraction of key parameters, such as the L3/L2 intensity ratio and peak positions, even in lower-resolution spectra. By providing robust oxidation state and hybridisation analysis under realistic imaging conditions, the fitting methodology offers a practical tool for redox quantification in beam-sensitive materials like NMC811. To evaluate redox changes and degradation in NMC811, two key quantitative EELS metrics were used, based on work by Hwang et al.:26 • L3/L2 Ratio: Reflects transition metal oxidation states. This ratio decreases with oxidation (due to changes in 3d electron occupancy) and increases with reduction. • ΔE Parameter: The energy separation between the oxygen pre-peak and main peak. A lower ΔE value suggests reduced metal-oxygen hybridisation, often associated with transition metal reduction or oxygen vacancy formation. The multi-Gaussian fitting approach addresses a fundamental challenge in EELS analysis: extracting quantitative information from spectra where instrumental broadening obscures oxidation-state-sensitive features. While the nominal energy resolution of non- monochromated systems (~1.5 eV FWHM) is often cited as a limitation, curve fitting methods can achieve sub-channel precision in peak position determination when applied to well- defined spectral features with adequate signal-to-noise ratio. This principle, established in spectroscopy more broadly and applied in EELS for peak position analysis, enables detection of chemical shifts smaller than the dispersion channel width.29,260 For the L3/L2 white-line ratio, which reflects 3d orbital occupancy, the fitting approach separates overlapping contributions 124 5.3 Quantitative EELS Mapping of Redox States in NMC811 and provides integrated intensities that would otherwise require visual deconvolution. Similarly, the oxygen ΔE parameter, derived from fitted pre-peak and main-peak positions, offers a measure of metal-oxygen hybridisation that is sensitive to oxidation state changes at the surface.260,261 By fitting rather than simply integrating spectral windows, the methodology produces more robust metrics that are less sensitive to energy calibration drift and background subtraction artifacts.130 This is particularly important for mapping applications, where pixel-to-pixel consistency enables meaningful spatial comparison of redox gradients. Reference spectra for different oxidation states were obtained for comparison. Ni2+ and Ni3+ spectra were acquired from NiO and LiNiO2, respectively. Spectra for Ni4+ were excluded due to the instability of highly delithiated compounds like LNO, which degrade rapidly under electron beam exposure or air contact and are spectrally similar to Ni3+, potentially leading to misinterpretation. Manganese reference values were obtained from standard specimens measured at SuperSTEM by Dr Demie Kepaptsoglou. While absolute peak positions may vary slightly between instruments, relative energy separations between oxidation states remain reliable indicators for oxidation state analysis. Given the beam sensitivity of Ni-rich layered oxides, this study prioritised specimen integrity by acquiring dose-efficient spectra and applying spectral fitting techniques. The multi- Gaussian fitting approach enabled reliable tracking of redox changes and hybridisation trends without introducing significant beam damage. Importantly, the methodology developed here is broadly applicable to similar studies, as the use of non-monochromated systems with moderate dispersion settings is common across many research laboratories. By offering a practical solution for extracting quantitative redox information from widely accessible EELS setups, this work provides a valuable tool for the broader battery research and electron microscopy communities. 5.3.3 Pre-processing: Thickness Normalisation and Background Removal A key objective of this study was to spatially resolve degradation in cycled NMC811 particles, particularly the formation of rock-salt (NiO-like) surface phases enriched in Ni2+. These phases are known to form at particle surfaces and grain boundaries under high voltage or prolonged cycling. To minimise thickness-related artefacts in the EELS signal, spectrum images were acquired from FIB-prepared sections with relatively uniform thickness profiles along the y axis. This consistent geometry enabled binning along one of the axes of the spectrum image (SI), e.g. the y-axis, which runs parallel to the particle surface and perpendicular to the degradation gradient, without sacrificing spatial resolution across depth. Along the x axis, there is a 5.3 Quantitative EELS Mapping of Redox States in NMC811 125 thickness gradient, and this is attributed to the difference in λ values between the layered NMC811 and rock-salt structure as discussed in Section 3.4.1. The lower t/λ value along the surface is indicative of a degraded region. In regions where signal intensity was low, such as the Mn and Co edges, binning improved the signal-to-noise ratio. Areas with t/λ < 0.5 (roughly 65 nm for NMC811) were selected to minimise plural scattering. Zero-loss peak alignment was used to correct for energy drift across the spectrum image. Figure 5.4: (a) HAADF STEM image of NMC811 cycled under the LowStart protocol with a selected spectrum image region highlighted in the green box. The corresponding maps of the spectrum image are shown in (b) for t/λ and (c) for thickness in nm. Figure 5.4a shows the selected region from a LowStart-cycled NMC811 particle, while Figure 5.4b and Figure 5.4c present the corresponding t/λ and absolute thickness maps. A clear thinning is observed near the surface, consistent with a wedge shape (due to crystal faceting), with relatively more intense evidence of oxygen loss and structural reconfiguration during the formation of rock-salt phases. Background removal is essential for accurate quantification of the L3/L2 white-line ratios and oxygen ΔE values. Standard power-law fitting, as implemented in HyperSpy, often fails to model the background correctly, particularly beneath the L2 edge and O K-edge main peak, leading to imprecise ratio calculations. This issue is illustrated in Figure 5.5, where the grey dashed line shows poor baseline fits. 126 5.3 Quantitative EELS Mapping of Redox States in NMC811 Figure 5.5: Standard background subtraction using PowerLaw functions in HyperSpy. (a) Oxygen K-edge and (b) Nickel L-edge spectra with poor baseline fitting (grey dashed line) To address this, a polynomial background subtraction method was developed, using element- specific pre- and post-edge fitting ranges and polynomial orders tailored to each edge. These parameters are summarised in Table 5.2. These parameters enable accurate background modelling by accounting for the complex slope of the signal under each edge, particularly for the oxygen K-edge, which required a fourth- order polynomial. These parameters function as follows: • Edge Range: The complete energy window containing the core-loss edge of interest • Pre-edge Fitting Range: The region before the edge onset used to sample the background signal • Post-edge Fitting Range: The region after the edge used to constrain the background model's trajectory • Polynomial Order: The complexity of the function used to model the background— higher orders accommodate more complex background profiles. Oxygen necessitated a fourth-order polynomial due to the particularly complex background profile beneath its K-edge features. O-K1 Pre-edge O-K2 Main edge (a) Ni-L3 Ni-L2 (b) 5.3 Quantitative EELS Mapping of Redox States in NMC811 127 Table 5.2: Optimised background fitting parameters for EELS analysis of NMC811 components Ni Mn Co O Edge Range / eV 820 - 900 590 - 690 720 - 820 450 - 590 Pre-edge Fitting / eV 790 - 850 580 - 635 720 - 770 460 - 510 Post-edge Fitting / eV 885 - 900 662 - 680 805 - 820 556 - 580 Polynomial 3 2 2 4 Figure 5.6: Background subtraction methodology for EELS analysis of NMC811 cathode materials. Panels show (a) Ni L-edge, (b) Mn L-edge, (c) O K-edge and (d) Co L-edge with original data (red), fitted background (blue dashed) constrained by pre-edge (light yellow) and post-edge (light orange) regions, and resulting background-removed signals (black). Coloured regions highlight the relevant L3 and L2 edges for transition metals and pre- peak/main peak structures for oxygen 128 5.3 Quantitative EELS Mapping of Redox States in NMC811 As shown in Figure 5.6, the fitted background (blue dashed line) closely follows the pre- and post-edge regions, resulting in clean background-subtracted signals (black lines) that clearly resolve the L3/L2 edges for transition metals and the pre-peak/main peak structure for oxygen. Background-corrected EELS data provides insights comparable to X-ray absorption spectroscopy (XAS) for analysing transition metal oxidation states and oxygen coordination environments in cathode materials125,141,262. This analytical equivalence enables direct correlation between our EELS findings and the established XAS literature on similar NMC cathodes, particularly when tracking changes in transition metal valence states and oxygen participation in redox processes during electrochemical cycling. While EELS offers significantly higher spatial resolution at the nanoscale compared to conventional XAS techniques, the spectral features after appropriate background removal exhibit equivalent characteristics. This similarity is especially pronounced in white-line intensity ratios of transition metal L-edges and pre-edge features in oxygen K-spectra, both critical indicators of electronic structure evolution. The complementary nature of these techniques strengthens our interpretation of local chemical changes occurring at cathode particle surfaces during different charging protocols. These preprocessing steps, thickness control and accurate background correction, were essential to ensure the reliability of the subsequent Gaussian fitting analysis, enabling precise mapping of redox gradients and oxygen bonding variations in cycled NMC811 cathodes. 5.3.4 Gaussian Fitting A key methodological contribution of this work is the development of a multi-Gaussian fitting algorithm to deconvolute overlapping spectral features in EELS, particularly where limited energy resolution prevents direct visual discrimination of adjacent oxidation states. This is critical for accurately resolving redox-active species in transition metals, where shifts between states such as Ni2+ and Ni3+ are often smaller than the resolution limit. Each spectral component is modelled using Gaussian functions defined as: 𝐺(𝑥; ℎ, 𝜎, 𝜇) = ℎ. exp (− (𝑥 − 𝜇)2 2𝜎2 ) Equation 20 where ℎ represents the peak amplitude, 𝜎 defines the peak width, and 𝜇 the peak centre position in energy space. 5.3 Quantitative EELS Mapping of Redox States in NMC811 129 Complex spectral features are modelled using linear combinations: 𝐺𝑡𝑤𝑜(𝑥) = 𝐺(𝑥; ℎ1, 𝜎1, 𝜇1) + 𝐺(𝑥; ℎ2, 𝜎2, 𝜇2) 𝐺𝑡ℎ𝑟𝑒𝑒(𝑥) = 𝐺(𝑥; ℎ1, 𝜎1, 𝜇1) + 𝐺(𝑥; ℎ2, 𝜎2, 𝜇2) + 𝐺(𝑥; ℎ3, 𝜎3, 𝜇3) Equation 21 Equation 22 where the numerical subscripts denote individual components. Figure 5.7: (a) Single Gaussian function defined by its peak height ℎ, centre position 𝜇, and standard deviation 𝜎, which controls the width of the curve. (b) Double Gaussian function composed of two individual Gaussian components: one with parameters ℎ1, 𝜇1 , 𝜎1 (green) and the other with ℎ2, 𝜇2 , 𝜎2 (blue). The sum of both components gives the total fitted peak 𝐺𝑡𝑤𝑜 (red) As illustrated in Figure 5.7, representing the spectrum as a sum of Gaussian components provides an effective route to separate overlapping contributions and to obtain continuous estimates of peak centre energies and integrated intensities. This is essential for producing spatial maps of redox-sensitive metrics where small shifts and weak shoulders would otherwise be obscured by noise or limited dispersion. Two fitting windows were defined for each feature (Figure 5.8): an ‘energy range’ that sets the overall fitting window, and a narrower ‘peak range’ that constrains the region in which the optimiser can assign the peak centre. This separation improves robustness by preventing the optimiser from locking onto neighbouring features or background residuals while still allowing the full line shape to be fitted. This algorithm allows for dynamic tracking, which is crucial when the Ni-L3 peak shifts from its typical position in layered bulk phase (Ni3+) to lower energy in surface rock-salt phase (Ni2+). (a) (b) 130 5.3 Quantitative EELS Mapping of Redox States in NMC811 Figure 5.8: Definition of energy range and peak range used in Gaussian fitting of the Ni L3 and L2 edges. The energy range (blue arrows) represents the full window selected to isolate the spectral feature, while the peak range (black arrows) defines a narrower fitting window (±2 eV) centred around the peak maximum to focus the fitting and avoid interference from neighbouring features. Energy and peak ranges were optimised for each spectral feature as shown in Table 5.3. The peak range varies depending on feature width to ensure fitting focuses on the most relevant spectral region. Rather than using fixed positions, the algorithm locates intensity maxima and adjusts fitting windows accordingly, adapting to local chemical shifts. Non-linear least-squares fitting was performed using a Trust Region Reflective algorithm with constraints to enforce physically meaningful solutions. Peak amplitudes were restricted to positive values, widths were bounded to avoid unrealistic peak shapes, and starting parameters were initialised to promote convergence across all pixels. These constraints reduce sensitivity to local minima and improve the stability of parameter maps in low-signal regions. A practical feature of the implementation is the automatic sorting of Gaussian components by centre energy. This ensured consistent peak assignment across all pixels, which is essential for mapping trends across surface-to-bulk gradients. The algorithm consistently labels the leftmost and rightmost peaks (e.g. g1 and g2), enabling reliable comparison of redox shifts and bonding changes between different regions of the particle. Peak Range 7 eV Energy Range Peak Range Energy Range Ni-L3 Ni-L2 5 eV 5.3 Quantitative EELS Mapping of Redox States in NMC811 131 5.3.5 Quantitative Metrics for Oxidation State Determination The fitted spectrum is reconstructed as the sum of all Gaussian components and the background, producing a composite curve that can be compared directly to the raw EELS data (Error! Reference source not found.). This reconstruction provides an internal check on fit q uality and ensures that derived metrics are based on the full fine structure rather than on single-channel peak maxima. The summation of Gaussian components, purple for oxygen peaks, orange for Mn, blue for Co, and green for Ni, produces a composite curve that fits the raw EELS spectra well, as shown in Error! Reference source not found.. For all transition metal L -edges and the oxygen pre-peak, two Gaussians were sufficient, while the O-K2 main peak required three. The code used for fitting automatically saves the results at each spatial position, facilitating later inspection and adjustment of fitting parameters, such as those listed in Table 5.3, should refinements be needed. Table 5.3: Spectral fitting constraints for core-loss edges in NMC811. The energy range defines the total span considered for fitting, and the peak range indicates the ±eV window centred around the detected maximum for refining the fit. ** denotes double gaussian and *** denotes triple gaussian used for fitting. Element Spectral Feature Energy Range (eV) Peak Range (eV) Ni ** L3 edge 852.0 - 862.0 ±7.0 Ni ** L2 edge 870.0 - 879.0 ±5.0 Mn ** L3 edge 638.0 - 648.0 ±8.0 Mn ** L2 edge 650.0 - 659.0 ±8.0 Co ** L3 edge 776.0 - 788.0 ±10.0 Co ** L2 edge 792.0 - 800.0 ±6.0 O ** K pre-peak (K1) 525.0 - 534.0 ±8.0 O *** K main peak (K2) 535.5 - 553.0 ±20.0 132 5.3 Quantitative EELS Mapping of Redox States in NMC811 From the fitted curves, two primary metrics were extracted: • The L3/L2 white-line ratio, calculated using Simpson’s rule for numerical integration, serves as an indicator of oxidation state, especially useful for overlapping or poorly resolved transition metal edges. For Ni, this ratio ranges from approximately 2.5-3.0 for Ni2+ and 1.8-2.3 for Ni3+, providing a reliable way to track redox evolution. • The oxygen ΔE parameter, calculated from the difference in fitted peak centre positions of the O-K1 and O-K2 components, reflects changes in metal-oxygen hybridisation and can indicate oxygen loss or transition metal reduction These parameters were mapped as a function of distance from the particle surface, enabling nanometre-scale resolution of redox gradients and degradation features. Compared to conventional area integration, this multi-Gaussian framework enables more precise extraction of electronic structure information, even in cases where spectral features are partially overlapped or energy-shifted due to local chemical changes. demonstrates the practical value of the multi-Gaussian fitting approach for extracting quantitative information from spectra acquired under realistic, non-monochromated conditions. The summed Gaussian curves closely reproduce the experimental data across all edges, validating the model assumptions. Critically, the decomposition reveals features that would be difficult or impossible to quantify through direct integration alone. For instance, the oxygen pre-peak (O-K1), which reflects metal-oxygen hybridisation and is sensitive to transition metal reduction, overlaps significantly with the rising edge of the main peak. Without fitting, attempts to define integration windows inevitably incorporate contributions from both features, introducing systematic errors that vary with oxidation state. Similarly, the Ni-L3 edge exhibits asymmetry arising from multiplet structure that, while not fully resolved at 1.5 eV resolution, manifests as a broadened shoulder. The two-Gaussian model captures this asymmetry, enabling more accurate intensity determination than symmetric integration would allow. The automatic saving of fitting parameters at each spatial position facilitates retrospective analysis and parameter optimisation, supporting the reproducibility that is essential for meaningful spatial mapping across degradation interfaces. 5.3 Quantitative EELS Mapping of Redox States in NMC811 133 Figure 5.9: Multi-Gaussian fitting of EELS spectra for key edges in NMC811. Double Gaussian fitting for (a) O-K1 pre-edge region, (c) Mn-L3, (d) Mn-L2, (e) Co-L3, (f) Co-L2, (g) Ni-L3, (h) Ni- L2 edges, and triple Gaussian fitting of (b) O-K2 main edge. Grey circles represent raw EELS data; black line indicate individual Gaussian components (g1, g2, g3), and their summed fit in their respective colours. Intensity values have been normalised to 1. (a) (b) (c) (d) (e) (f) (g) (h) 134 5.4 EELS Results 5.4 EELS Results 5.4.1 Spatial Evolution of Transition Metal Redox and Oxygen Chemistry Figure 5.10 presents background-subtracted, Gaussian-fitted EELS spectra for the O K-edge and transition metal L-edges (Ni, Mn, Co) from pristine, LowStart-cycled, and LowEnd-cycled NMC811 samples. Each dataset compares spectra from the surface (lighter shade) and bulk (darker shade) regions. Figure 5.10: Background-subtracted EELS spectra of LowEnd, LowStart, and pristine NMC811 samples at both surface (lighter shade) and bulk (darker shade) regions. Spectra for (a) O K-edge, (b) Ni L-edge, (c) Mn L-edge, and (d) Co L-edge 5.4 EELS Results 135 In the oxygen K-edge (Figure 5.10a), the pre-peak intensity is notably suppressed at the surface of both cycled samples, indicating reduced O 2p-TM 3d hybridisation, likely caused by oxygen vacancy formation and transition metal reduction. The ΔE parameter, reflecting the energy separation between the pre-peak and main peak, also contracts at the surface, particularly in the LowStart sample, consistent with weakened metal-oxygen bonding. In contrast, the pristine sample exhibits uniform pre-peak intensity and ΔE values across surface and bulk, confirming an undisturbed electronic structure. The Ni L3 peak in both cycled conditions shifts to lower energy at the surface, consistent with Ni3+ → Ni2+ reduction, and is most pronounced in LowStart-cyced samples, whose surface closely resembles that of NiO. Mn and Co edges exhibit subtler surface shifts, with weaker signals due to their lower concentration in NMC811 (Ni:Mn:Co ≈ 8:1:1), but still suggest partial reduction, particularly in the cycled samples. It is worth noting that the weaker signals of Mn and Co result in increased noise, which can affect both the L3/L2 ratio and the accuracy of peak centre determination from Gaussian fitting. In contrast, the pristine sample showed no significant changes in any of the transition metal edges, confirming the absence of surface degradation and validating that the observed modifications in cycled samples arise from electrochemical processes rather than artefacts of sample preparation. 5.4.2 Protocol-Dependent Degradation Gradients To quantitatively assess redox gradients, Figure 5.12 maps the spatial evolution of L3/L2 ratios for Ni, Mn, and Co, along with the oxygen ΔE parameter, for all three sample types. Reference values for assigning oxidation states are summarised in Table 5.4. Ni2+ and Ni3+ references were obtained from NiO and LiNiO2, respectively. Ni4+ reference spectra, although collected from delithiated LNO, were excluded due to spectral similarity to Ni3+. Mn reference spectra were measured by Dr. Demie Kepaptsoglou at SuperSTEM. While minor instrument-induced peak shifts may exist, the relative differences between oxidation states remain valid. Literature values were also used to validate these references. The reference L3/L2 ratios in parentheses represent literature values that were used to validate our measurements. Figure 5.11 and Figure 5.12 present the same EELS dataset, one particle each from LowStart, LowEnd, and pristine conditions, visualised in two complementary ways. Figure 5.11 enables direct comparison of each spectroscopic parameter (Ni, Co, Mn L3/L2 ratios and oxygen ΔE) across the three protocols, while Figure 5.12 displays all parameters together for each sample, revealing how elemental redox changes evolve relative to one another within a single depth profile. The shaded halos in both figures represent error estimates based on the energy dispersion channel (±0.3 eV) used during EELS acquisition. 136 5.4 EELS Results Table 5.4: Reference L3/L2 ratios and peak positions for key oxidation states. Ni references acquired in-house; Mn from SuperSTEM; Co and additional Mn values adapted from literature,263 denoted as * for Mn to distinguish between SuperSTEM and reference values. Oxidation State L3/L2 Ratio Ni L3 Center (eV) Ni L2 Center (eV) Ni2+ 3.6 - 4.3 853.6 - 853.8 871.4 - 871.6 Ni3+ 2.9 - 3.7 853.8 - 854.3 871.8 - 872.5 Mn2+ 3.2 - 3.8 4.2 - 4.4 * 639.4 - 640.0 651.7 - 652.1 Mn3+ 2.4 - 2.6 2.2 - 2.3 * 639.4 - 640.4 650.5 - 651.7 Mn4+ 1.8 - 1.9 1.8 - 2.1 * 641.3 - 641.6 651.8 - 652.2 Co2+ 4.7 - 5.3 Co2.67+ 3.2 - 3.4 Co3+ 2.7 Co4+ 1.9 - 2.2 Among the four parameters, Ni and oxygen ΔE exhibit the clearest trends. This is attributable both to their substantially higher signal-to-noise ratios and to their direct involvement in the surface phase transformation from layered to rock-salt structure. In LiNi0.8Mn0.1Co0.1O2, Ni constitutes 80% of the transition metal content, whereas Mn and Co each account for only 10%, resulting in correspondingly weaker L-edge signals. Oxygen, present in a 2:1 stoichiometric ratio relative to the combined transition metals, also benefits from strong signal intensity. Critically, the layered-to-rock-salt transformation is driven primarily by Ni3+→Ni2+ reduction and concurrent oxygen loss, making these two parameters the most sensitive indicators of surface degradation. Manganese and cobalt, by contrast, remain largely redox-inactive within the cycling voltage window and are therefore expected to show minimal systematic variation. 5.4 EELS Results 137 Even with the conservative error bounds shown, several systematic trends in Ni and oxygen ΔE remain distinguishable from measurement noise. For Ni, the LowStart sample shows an elevated L3/L2 ratio (~3.7) at the surface that decreases with depth before plateauing at bulk values ( ~ 3.1) beyond approximately 20 nm. This gradient is consistent with Ni3+→Ni2+ reduction confined to a surface reconstruction layer. The LowEnd sample exhibits a similar but narrower gradient, with the ratio stabilising within 8 nm of the surface, indicating a thinner chemically degraded region. The pristine sample shows no systematic depth dependence, confirming the absence of surface reduction prior to cycling. The oxygen ΔE parameter mirrors these trends: LowStart shows a gradual rise that plateaus only beyond 20 nm, while LowEnd increases more sharply within the first ~ 8 nm before reaching bulk values (~12 eV). The lower ΔE values near the surface reflect diminished metal- oxygen hybridisation, consistent with oxygen vacancy formation and rock-salt-like phases. The pristine sample maintains relatively constant ΔE throughout the measured depth. For cobalt and manganese, the L3/L2 profiles show considerable scatter without clear systematic trends, attributed to lower signal intensity from these minority elements and their limited participation in the redox processes within this voltage window. Manganese is expected to remain predominantly in the Mn4+ state, serving as a structural stabiliser, while Co undergoes only minor valence changes compared to Ni.26 Figure 5.12 presents the same data as an intra-sample comparison, displaying all parameters together for each protocol. This visualisation reinforces that the spectroscopic changes originate predominantly from nickel reduction and associated oxygen ΔE decrease, while Mn and Co traces exhibit fluctuations that overlap substantially with their error bands. The pristine panel confirms baseline behaviour, with oxygen ΔE relatively flat (11–12 eV) and transition metal ratios fluctuating around expected bulk oxidation states (Ni3+, Co3+, Mn4+); minor variations likely reflect local compositional heterogeneity rather than degradation. 138 5.4 EELS Results Figure 5.11: Comparative depth profiles of L3/L2 area ratios and oxygen ΔE parameter across different cycling protocols. Fluctuations in Mn and Co trends are attributed to low signal-to-noise. Shaded halos represent error bars based on the dispersion channel during EELS acquisition. 5.4 EELS Results 139 Figure 5.12: Depth profiles of L3/L2 area ratios for Ni, Co, and Mn (left axis) and oxygen ΔE parameter (right axis) for NMC811 samples cycled under LowStart, LowEnd, and pristine conditions. Fluctuations in Mn and Co trends are attributed to low signal-to-noise. Shaded halos represent error bars (±0.3 eV) based on the dispersion channel during EELS acquisition 140 5.4 EELS Results 5.4.3 Peak Centre Position Analysis While white-line ratios (L3/L2) provide reliable indicators of oxidation state, shifts in the L3 and L2 peak centre positions offer an independent and complementary method to assess local changes in electronic structure. These shifts are particularly valuable in complex systems like NMC811, where overlapping features and noise can affect intensity-based metrics. The Ni L3 centre, in particular, is highly sensitive to oxidation state, typically shifting to lower energies as nickel is reduced from Ni3+ to Ni2+. Figure 5.13 presents the spatial evolution of peak centre positions for the Ni, Mn and Co L2,3 edges, and the oxygen pre-peak and main peak across a LowStart-cycled NMC811 sample. These positions were extracted from Gaussian-fitted spectra at each pixel across the surface- to-bulk gradient. In the Ni L3 edge (Figure 5.13a), a clear redshift is observed at the surface, with the peak centre decreasing from 856.0 eV in the bulk to 855.2 eV at the surface. This shift is consistent with a transition from Ni3+ to Ni2+, supporting earlier L3/L2 ratio trends and confirming the formation of a reduced rock-salt-like surface layer. The transition occurs over approximately 20 nm, matching the redox gradient width previously identified in Figure 5.12a. The oxygen pre-peak shifts from ~533.0 to ~529.5 eV and the main peak from ~538.5 eV to 542.0 eV consistent with oxygen loss and rock-salt phase formation. Shifts in the Mn and Co L3 peak centres (Figure 5.13c and Figure 5.13f) are less pronounced, reflecting their relatively stable oxidation states during cycling. Mn and Co also show redshifts in L3 peaks, with their L2 peaks relatively stable suggesting partial reduction. The uncertainty treatment in Figure 5.13 affects the strength of the protocol comparison in two ways. First, it provides a threshold for distinguishing genuine near-surface shifts (which persist across multiple adjacent pixels and exceed the error envelope) from isolated pixels that may be influenced by low counts or local background variations. Second, it supports the interpretation that the observed gradients are systematic spatial trends rather than random fitting scatter, strengthening the link between spatially resolved redox evolution and the macroscopic electrochemical behaviour. Although Figure 5.13 presents peak centre analysis for the LowStart sample, equivalent results for LowEnd and pristine samples are provided in Figure A1 and A2 in the Appendix, respectively. In the LowEnd case, Ni L3 shifts were less pronounced and confined to a narrower ~10 nm surface region, consistent with reduced surface reconstruction. The pristine sample shows no significant variation in any edge positions. 5.4 EELS Results 141 Figure 5.13: Spatial evolution of EELS peak centre positions from the particle surface towards the bulk across a LowStart-cycled NMC811 particle. (a-b) Ni L3 and L2 edges, (c-d) Co L3 and L2 edges, (e-f) Mn L3 and L2 edges, and (g-h) O K pre-peak and main peak positions. Error bars represent the conservative uncertainty associated with the spectral dispersion (0.3 eV per channel) and fit stability, as described in the text. The maps visualise the systematic near-surface shift relative to the particle core that underpins the protocol comparison. To summarise the trends across all three samples, Table 5.5 presents the full range of measured L3/L2 ratios, ΔE values, and peak centre positions. These data confirm the more extensive redox and structural changes under the LowStart protocol, while highlighting the mitigated degradation in LowEnd. It is worth noting that the measured peak positions for Mn differ somewhat from the reference values from SuperSTEM in Table 5.4, likely due to instrumental differences between our setup and the reference measurement systems. Despite these peak shifts, the L3/L2 ratio remains a more reliable metric for oxidation state determination, as it is less sensitive to absolute energy calibration and more indicative of the fundamental electronic structure. Importantly, our (a) (b) (c) (d) (e) (f) (g) (h) 142 5.4 EELS Results fitting methodology successfully produced well-fitted Mn data without requiring parameter adjustment of the Gaussian fitting parameters established in Table 5.3, demonstrating the robustness and versatility of our approach across different elements and experimental conditions. This consistency is particularly valuable when analysing complex, heterogeneous samples where multiple elements must be tracked simultaneously across degradation gradients. Table 5.5: Summary of L3/L2 ratios, ΔE, and peak centre positions for pristine, LowStart, and LowEnd samples. Mn and Co values fluctuate due to low signal. ΔE indicates O-K pre/main peak separation; degradation zone depth is estimated from combined redox and peak shift profiles. Pristine LowStart LowEnd Ni L3/L2 2.7 - 3.2 3.3 - 2.8 3.7 - 3.1 Mn L3/L2 1.3 - 1.7 1.8 - 2.2 2.1 - 2.7 Co L3/L2 1.8 - 2.9 2.3 - 3.5 1.6 - 2.6 O ΔE parameter 11.5 6.0 - 12.0 5.0 - 12.5 Ni L3 center 856.6 - 857.2 855.2 - 856.0 855.7 - 857.0 Ni L2 center 873.3 - 874.7 873.2 - 873.8 873.5 - 874.8 Mn L3 center ~ 645.0, fluctuates 643.0 - 643.5 642.0 - 644.2 Mn L2 center ~ 655.0, fluctuates ~ 654.3, fluctuates ~ 654.0, fluctuates Co L3 center ~ 782.5, fluctuates 781.4 - 782.0 781.5 - 783.0 Co L2 center ~ 797.5, fluctuates 796.4 - 796.6 ~ 797.0, fluctuates O pre-peak ~ 529.6 530.0 - 529.0 533.0 - 529.5 O main peak ~ 540.0, fluctuates 540.2 - 540.8 538.5 - 542.0 Depth of degradation zone None ~20 nm ~10 nm 5.4 EELS Results 143 5.4.4 Direct Structural Visualisation of Phase Transformations To complement the spectroscopic EELS analysis and validate the proposed degradation mechanisms, high-resolution HAADF-STEM imaging was carried out on cycled and pristine NMC811 particles. These structural investigations offer direct crystallographic evidence of the phase transformations and microstructural changes that occur during electrochemical cycling. Figure 5.14: Atomic-resolution imaging of structural phases in LowEnd NMC811. (a) HAADF- STEM image showing the RSL-spinel-layered transition. (b-d) High-magnification images overlaid with structural models of each phase. Atom colours: yellow = Li, blue = Transition metal, red = O. Figure 5.14 presents atomic-resolution STEM images illustrating a clear structural progression from surface to bulk within a cycled NMC811 particle. In Figure 5.14a, three distinct regions are identified: a rock-salt-like (RSL) phase at the surface, a transitional spinel-like region, and a well-ordered layered structure in the particle core. This crystallographic sequence mirrors the oxidation state gradient captured in EELS L3/L2 profiles. The high-magnification images in Figure 5.14b-d overlaid with structural models, confirm these assignments. The layered region (Figure 5.14b) exhibits clear transition metal (TM) layers separated by lithium channels, consistent with predominantly Ni3+ content. The spinel region (Figure 5.14c) reveals partial cation disorder, indicative of a mixed Ni2+/Ni3+ environment. In contrast, the RSL region (Figure Figure 5.14d) lacks lithium channels and adopts a cubic structure, consistent with Ni2+ dominance at degraded surfaces. 144 5.4 EELS Results Figure 5.15: Cross sectional HAADF-STEM overview images of NMC811 particles after cycling under (a) LowStart, (b) LowEnd, and (c) pristine conditions. Note the extensive cracking in LowEnd versus the more intact morphology of LowStart Protocol-dependent differences in microstructural damage are shown in Figure 5.15. The LowEnd sample (Figure 5.15b) exhibits widespread intergranular cracking throughout the secondary particle, indicative of pronounced mechanical degradation. In contrast, the LowStart sample (Figure 5.15a) shows a more intact morphology with fewer visible cracks, suggesting better structural cohesion. The pristine sample (Figure 5.15c) displays no observable cracking, retaining a dense and uniform structure. These differences indicate that while LowEnd induces more mechanical damage, LowStart undergoes more extensive chemical degradation, as evidenced by deeper redox gradients and thicker surface- transformed layers in EELS analysis. Although the LowEnd sample exhibits more pronounced intergranular cracking in the STEM images, this mechanical damage did not translate into poorer electrochemical performance in these cells. Instead, LowEnd maintained higher capacity retention, consistent with its reduced high-voltage dwell. This points to a partial decoupling between mechanical fracture and the chemical degradation quantified by EELS, that limiting high-voltage exposure can suppress surface reconstruction and impedance growth even when cracking is present. This observation reinforces the value of protocol optimisation as a practical route to mitigate the most performance-limiting degradation pathways. 203 Intragranular cracks also serve as local initiation sites for phase transformation. Figure 5.16 provides high-resolution imaging near a crack face in a LowStart-cycled particle. The red box in Figure 5.16a highlights the region selected for detailed analysis. The zoomed-in image (Figure 5.16b ) shows clear lattice fringes, and the line profile in Figure 5.16c captures the local variation in interplanar spacing. Near the crack, a d-spacing of ~2.54 Å confirms the formation of a rock-salt-like phase. Moving away from the crack, alternating ~ 1.92 Å and ~ 3.02 Å (b) (a) LowStart LowEnd (c) Pristine 5.4 EELS Results 145 spacings are observed, characteristic of a distorted spinel structure. Further into the particle, the lattice spacing increases to ~4.96 Å, consistent with the pristine layered structure.53 These measurements demonstrate a localised crystallographic transition from layered → spinel → RSL across the crack interface. The spatial variation aligns well with EELS redox gradients, highlighting that cracks not only compromise mechanical integrity but also serve as pathways for chemical degradation, accelerating the formation of electrochemically inactive phases. These structural observations provide critical validation of the redox gradients identified by EELS, confirming that electrochemical degradation in NMC811 involves both phase transformation and microstructural evolution. Together, they underscore the importance of controlling both surface reconstruction and mechanical integrity when designing cycling protocols to preserve cathode performance. Figure 5.16: HAADF-STEM analysis of crack-induced degradation for LowStart sample. Line profiles show a transition from rock-salt (~2.54 Å) near the crack to layered structure (~4.96 Å) deeper into the particle. 146 5.5 Integrating Spectroscopic and Electrochemical Insights 5.5 Integrating Spectroscopic and Electrochemical Insights The comprehensive spectroscopic and electrochemical characterisation of NMC811 cathodes under different charging protocols reveals a coherent picture of degradation pathways across multiple scales. Electrochemical performance data demonstrated that the LowEnd protocol achieved the most stable long-term capacity retention and the lowest impedance growth, while LowStart, despite exhibiting fewer mechanical fractures, suffered from significantly more severe surface chemical degradation. These trends are not only mirrored but mechanistically explained through EELS analysis. Electron Energy Loss Spectroscopy enabled high-resolution mapping of redox gradients and surface transformation layers. In LowStart samples, a ~ 20 nm thick surface reconstruction layer, dominated by reduced Ni2+ states and diminished metal-oxygen hybridisation, was consistently observed. This degraded layer, comprising rock-salt and spinel phases, is electrochemically inactive and presents a barrier to lithium-ion transport. Its formation aligns closely with the prolonged CV phase inherent to the LowStart protocol, where the high terminal current just before the constant-voltage stage results in extended time at elevated potentials. This extended residence at 4.2 V likely promotes electrolyte oxidation, oxygen release, and transition metal reduction, accelerating the formation of this insulating surface layer. In contrast, LowEnd samples exhibited a much thinner (~ 10 nm) degradation zone, with redox states and hybridisation more closely resembling the pristine material. Despite exhibiting more mechanical cracking attributed, these samples maintained layered structures and electrochemical activity.70,203,264 This aligns with the electrochemical observation of a shortened CV phase, due to the declining current ramp toward the end of the charge cycle. The reduced time spent at high voltage appears to suppress parasitic surface reactions, thereby preserving electrochemical performance. Critically, the spectroscopic parameters, namely the L3/L2 white-line ratios and oxygen ΔE shifts correlate closely with electrochemical impedance trends. Higher L3/L2 ratios (Ni2+-rich) and reduced ΔE values (weakened metal-oxygen bonding) were consistently associated with increased charge transfer resistance and poorer capacity retention, as seen in the LowStart cells. This reinforces the conclusion that surface chemical degradation, not bulk mechanical damage, is the dominant limiting factor in long-term cycling under high-voltage conditions. Taken together, these results underscore the importance of charging protocols that reduce high-voltage dwell time, particularly within the CV phase. The integrated EELS- electrochemical analysis presented here offers a powerful approach to deciphering the interplay between atomic-scale degradation and macroscopic cell performance, guiding future strategies for designing more durable, high-energy-density cathode materials. 5.6 Conclusion 147 5.6 Conclusion This chapter demonstrates how spatially resolved EELS, combined with constrained multi-Gaussian peak fitting, can quantify redox evolution and oxygen-hybridisation changes in cycled NMC811 with nanometre-scale resolution. The fitting approach enables weak and overlapping spectral features to be separated consistently across spectrum images acquired under finite energy resolution, providing robust maps of L3/L2 and O-K metrics. Across the three charging protocols, the EELS-derived trends were consistent with the electrochemical ranking: LowStart showed the strongest near-surface changes indicative of increased chemical degradation, whereas LowEnd exhibited more stable near-surface electronic structure under reduced high-SoC dwell. Notably, LowEnd samples displayed more extensive mechanical cracking yet superior capacity retention, demonstrating a partial decoupling between mechanical fracture and chemical degradation. A central methodological contribution is the demonstration that multi-Gaussian fitting provides a practical route to quantitative EELS analysis on non-monochromated microscopes. By enabling reliable extraction of L3/L2 ratios and peak positions from spectra with ~1.5 eV resolution, the approach bridges the gap between high-throughput data acquisition and meaningful oxidation state quantification. The methodology proved robust across datasets exhibiting energy shifts of up to 4 eV, demonstrating generalisability beyond a single instrumental setup. This accessibility is significant: non-monochromated systems remain the workhorses of most electron microscopy facilities, and demonstrating that quantitative redox mapping is achievable without specialised equipment lowers the barrier to systematic EELS characterisation of battery materials. While this approach does not yield fully quantitative results grounded in quantum mechanical calculations such as Hartree-Slater265–267 or multiplet268 analysis, it offers a computationally efficient alternative for routine and high-throughput studies. In contrast to simulation-driven interpretation, the fitting routine developed here enables rapid, consistent comparison of oxidation-state-sensitive descriptors across large, beam-sensitive datasets. The analysis code and fitting scripts have been made publicly available as an open-source repository on GitHub269 to support reproducibility and enable broader adoption by the electron microscopy community. Designed with accessibility in mind, users need only define the background fitting regions and energy windows, followed by the Gaussian peak fitting parameters for their core-loss edges of interest. While this study focused on NMC811, the fitting routine is broadly applicable and can be used for any elemental edges in EELS datasets with overlapping spectral features, as long as they have sufficient signal-to-noise ratio. Once the parameters are set, the entire fitting process, including automated plotting of L3/L2 ratios, 148 5.6 Conclusion ΔE trends, and Gaussian peak shifts, can be completed in under five minutes per dataset, offering a fast, intuitive workflow for the broader microscopy community. Overall, the chapter shows that combining advanced spectroscopy with a carefully constrained fitting strategy can translate complex EELS fine structure into quantitative, spatially resolved descriptors of degradation, strengthening the mechanistic interpretation of protocol effects in Ni-rich cathodes. Chapter 6 149 Chapter 6 Classification of Nickel Oxidation States in NMC811 via EELS and Convolutional Neural Networks Building on the previous chapter’s multi-Gaussian fitting analysis of Ni L-edge and O K-edge features, this chapter introduces a convolutional neural network (CNN) approach for classifying Ni oxidation states from EELS spectra, using the same NMC811 cathode data. In this chapter we establish that traditional methods such as L3/L2 white-line ratios reveal redox gradients but lack the energy resolution to reliably distinguish between closely spaced states like Ni2+ and Ni3+ due to peak overlap and subtle spectral differences. Our CNN model, trained specifically on reference spectra of the Ni L3 edge with energy-shift augmentation, achieves 98.8% validation accuracy and significantly outperforms traditional analytical techniques. This enables pixel-by-pixel oxidation state mapping at nanometer-scale resolution, with potential application to high-resolution datasets given sufficient signal-to-noise ratio. In theory the data analysis could be applied to high res datasets too (as long as S/N is sufficient) When applied to NMC811 samples cycled under different protocols, the CNN identified distinct surface degradation patterns: a ~20 nm Ni2+-rich layer for the LowStart protocol and a thinner (~10 nm) reduction zone for LowEnd, correlating with differences in electrochemical performance. This approach bridges local spectroscopic analysis with macroscopic cell behaviour, demonstrating that surface redox degradation plays a more critical role than mechanical fracturing in limiting long-term performance. Finally, we explore the transferability of this model to X-ray absorption spectroscopy (XAS), establishing a framework for unified, multi-technique oxidation state analysis in battery research. 150 6.1 Introduction: Challenges in Direct Oxidation State Classification Acknowledgement: The electrochemical cycling of NMC811 materials used in this chapter was carried out by Dr Alexander Dimitrijevic at University College London. Helpful discussions on support vector machines with Dr Daniel del Pozo Bueno at Universidad de Barcelona sparked my interest in applying machine learning to EELS data. Mr Kelvin Chan, software engineer at Amazon, contributed by building the CNN architecture used in this study. Their contributions were instrumental in enabling the development of a machine learning-based approach for classification of Nickel ion from EELS data. All remaining aspects of this work, including dataset preparation, data preprocessing, training strategy design, model evaluation, EELS analysis, result interpretation, and figure preparation, were conducted by the main author. 6.1 Introduction: Challenges in Direct Oxidation State Classification Understanding redox evolution in Ni-rich layered cathodes such as NMC811 is crucial for interpreting degradation mechanisms and improving long-term battery performance.270 In the previous chapter, multi-Gaussian fitting of Ni L-edge and O K-edge EELS spectra enabled the extraction of L3/L2 white-line ratios and oxygen peak separations (ΔE), providing initial insight into spatial oxidation state gradients. These metrics revealed a general trend of Ni reduction from particle surfaces to the bulk. Near the surface, higher L3/L2 ratios (3.5 - 4.0) and smaller ΔE values were observed, consistent with Ni2+ formation and weakened metal-oxygen bonding. Deeper into the particle, lower L3/L2 ratios (~3.0) and larger ΔE indicated a return to Ni3+-dominated, layered-like bulk environments. However, while these indicators captured redox trends, they could not reliably identify which Ni oxidation state, particularly Ni2+ versus Ni3+, dominated at each spatial location. The Ni L2,3 edges for these states are separated by only ~ 1 - 2 eV, and their spectral features overlap significantly, forming a single broadened white-line envelope. As a result, intensity ratios and peak shifts function as approximate indicators of valence but lack the specificity to resolve mixed states or subtle transitions. It is experimentally and theoretically clear that the white- line ratio has no strict one to one correspondence with the oxidation state.271,272 Intermediate L3/L2 ratios, for example, could arise from a true Ni2+/Ni3+ mixture or from experimental factors such as sample thickness or background subtraction artefacts. To address this limitation, we turned to machine learning. An initial investigation using support vector machines (SVMs) highlighted the potential, but also the constraints, of traditional classification models. SVMs have previously been used to successfully classify Mn L2,3-edge spectra into Mn2+, Mn3+, and Mn4+, where the oxidation states are more clearly resolved140. However, when applied to the Ni energy loss near edge structure (ELNES), the subtlety of spectral differences and greater peak overlap limited the SVM’s ability to 6.1 Introduction: Challenges in Direct Oxidation State Classification 151 distinguish oxidation states effectively. These early results underscored the need for a more flexible, data-driven model, prompting the shift toward deep learning with convolutional neural networks (CNNs). CNNs are particularly well-suited for pattern recognition in complex, high-dimensional datasets such as EELS spectra141,273,274. Inspired by the human visual system, CNNs can learn hierarchical spectral features, starting with local slopes, edges, and intensities, and combining them into high-level representations of spectral shape. Crucially, they can learn directly from data without relying on hand-crafted rules or thresholds. In the context of Ni L-edge spectra, a CNN can be trained to recognise differences in peak width, symmetry, and energy position that correlate with specific oxidation states. While CNNs have also been applied to Mn L2,3-edge classification141, even in the presence of energy drift, no such model has yet been developed for Ni oxidation states from EELS. This likely reflects the greater challenge Ni presents: compared to Mn, Ni states exhibit smaller energy separations and more heavily overlapped features. Nonetheless, qualitative interpretation by experienced microscopists often proves effective, for example, broader, lower-energy L3 peaks are commonly associated with Ni2+. CNNs offer a way to replicate and scale this kind of expert judgement, enabling more definitive, high- resolution oxidation state maps. In this chapter, we: 1. Examine the limitations of conventional analytical and early machine learning methods for oxidation state classification; 2. Develop and optimise a CNN architecture specifically tailored to the Ni L-edge EELS spectra; 3. Apply the trained model to pixel-level classification of Ni oxidation states in NMC811 samples; 4. Compare CNN-derived maps to traditional results and correlate findings with electrochemical performance under different cycling protocols; and 5. Explore the potential for applying CNN-based EELS models to other spectroscopic techniques, such as X-ray absorption spectroscopy (XAS), to enable unified oxidation state analysis across modalities. 152 6.2 Challenges in Training Data Preparation for Machine Learning 6.2 Challenges in Training Data Preparation for Machine Learning A critical first step in applying deep learning is gathering reliable training and test data with known labels. In this study, the training data, derived from reference spectra representing known Ni2+ and Ni3+ states, remain fixed throughout the analysis and are used to teach the model to recognise characteristic spectral features of each oxidation state. In contrast, the test data are consistently drawn from three experimental samples: LowStart and LowEnd, which underwent different electrochemical cycling protocols, and a pristine uncycled sample. These datasets are used throughout this chapter to evaluate the model's performance and interpret the resulting oxidation state maps. Before implementing machine learning classification, we must first address fundamental challenges in preparing reliable training data for nickel oxidation state identification. 6.2.1 Limitations of the L3/L2 ratio approach By analysing the L3/L2 ratio of reference spectra, we identified significant spectral ambiguities that could potentially undermine classification accuracy. As shown in Figure 6.1, the L3/L2 ratio distributions for different nickel oxidation states from reference samples exhibit substantial overlap. Figure 6.1: Histograms showing the distribution of L3/L2 ratios measured from reference compounds: NiO (Ni2+), LiNiO2 (Ni3+), and delithiated LiNiO2 (Ni4+). Note the significant overlap between Ni3+ and Ni4+, and partial overlap between Ni2+ and Ni3+. Our analysis of reference compounds revealed that while Ni2+ typically exhibits higher L3/L2 ratios in the range of 3.6-4.3, there is partial overlap with Ni3+ in the range of 2.9-3.7. More problematically, the L3/L2 ratios for Ni4+ completely overlap with those of Ni3+, making these (a) (b) 6.2 Challenges in Training Data Preparation for Machine Learning 153 states virtually indistinguishable using this metric alone. This overlap is caused by the inherently similar spectra between Ni3+ and Ni4+, and as discussed in the previous chapter, the highly unstable Ni4+ state may reduce during sample preparation. Since the samples investigated are in the discharged state, the majority of the nickel present is expected to exist as Ni2+ or Ni3+, allowing us to focus our classification efforts on these two oxidation states. The comparison of L3/L2 ratios across our two test conditions, LowStart and LowEnd samples, yielded inconsistent results. The zones selected for different oxidation states were based on the mean ± standard deviation from the histogram plots. Part of the overlapped region (ratios between 3.31 and 3.41) considered a 'mixed region,' while Ni2+ and Ni3+ were defined by ratios above 3.41 and below 3.31, respectively. Figure 6.2: Classification maps based on L3/L2 ratios for (a) unbinned LowStart sample, (b) binned LowStart sample, (c) unbinned LowEnd sample, and (d) binned LowEnd sample. Blue regions represent Ni2+ (L3/L2 > 3.41), purple indicates mixed valence (3.31-3.41), and orange shows Ni3+ (L3/L2 < 3.31). The x-axis represents distance from the particle surface in nanometers. 154 6.2 Challenges in Training Data Preparation for Machine Learning Figure 6.2 shows attempts at classification using L3/L2 ratios, comparing both unbinned and binned datasets. Binning refers to the process of averaging neighbouring pixels or spectra to improve the signal-to-noise ratio in regions with low electron counts, particularly important for elements with weaker signals like Mn and Co. Nickel, by contrast, had sufficient signal intensity to permit unbinned analysis. Figure 6.2a and Figure 6.2c represent the original unbinned Ni dataset, Figure 6.2b and Figure 6.2d show the binned version, which is more consistent with the redox analysis approach used for other transition metals. The binned dataset applies a 10× binning perpendicular to the particle surface, identical to the procedure used in the Gaussian fitting analysis from the previous chapter, thus enabling a direct methodological comparison. In the binned version of the LowStart sample (Figure 6.2b), a strong indication of a Ni2+-rich layer within the first ~20 nm from the surface was observed, followed by a mixed valence zone. However, for the LowEnd sample (Figure 6.2d), the classification remained inconclusive. The unreliability of L3/L2 ratios for direct classification stems from several factors: 1. Background subtraction sensitivity: Small variations in background modelling can significantly alter the calculated L3/L2 ratios, particularly when the signal-to-noise ratio is low.130 2. Sample thickness and plural scattering: Thicker regions introduce additional scattering events that broaden and distort the spectral features which influence the apparent intensity ratios.130,272,275 3. Integration range dependency: The choice of energy windows for calculating peak areas introduces subjectivity and can affect the resulting ratios, especially since each EELS acquisition may have slight energy shifts requiring individual adjustment of integration windows.130 4. Mixed valence states: In real samples, nickel ions can exist in mixed valence states rather than discrete oxidation levels, further complicating the analysis. 6.2.2 Modification to feature selection Given the limitations of the L3/L2 ratio approach, we examined the spectral features themselves to identify more reliable discriminative characteristics. Figure 6.3 shows a comparison of reference spectra from Ni2+ and Ni3+ sources. The analysis revealed several key differences between the oxidation states: • The Ni2+ L3 peaks are sharper, more symmetric, and positioned at lower energy (~853.7 eV) • The Ni3+ L3 peaks are broader and shifted to higher energy (~854.2 eV) 6.2 Challenges in Training Data Preparation for Machine Learning 155 • The Ni2+ reference spectra show better consistency and less noise than the Ni3+ spectra • The L2 edge exhibits minimal shape variation between oxidation states, with differences primarily in relative intensity (Ni3+ has a higher L2 intensity than Ni2+, leading to lower L3/L2 ratios) • The L2 edge shows significantly more noise than the L3 edge in both oxidation state reference sam Figure 6.3: Comparison of reference spectra for Ni oxidation states. (a) Mean reference spectra for Ni2+ (blue) and Ni3+ (orange) showing distinct differences in L3 and L2 peak shapes. (b) Ten Ni3+ reference spectra showing consistency in L3 peak shape despite some noise in the L2 region. (c) Ten Ni2+ reference spectra showing sharper, more symmetric L3 peaks compared to Ni3+. 156 6.2 Challenges in Training Data Preparation for Machine Learning Initial attempts to train a CNN model using the entire Ni-L2,3 spectral range yielded inconclusive results, with classifications showing either a majority of pixels as Ni2+ or very random/scattered pixel classifications. The classification results of using both Ni L2 and L3 peaks are shown in Figure A6 in the Appendix. This was likely due to the noisy L2 region diluting the more distinctive features of the L3 edge. To improve classification accuracy, we modified our feature selection to focus solely on the L3 edge. By cropping the reference data to the 840 - 867 eV range, we concentrated on the spectral region with the highest signal-to- noise ratio and most pronounced differences between oxidation states. 6.2.3 L3 Peak Area Analysis as an Alternative Approach Prior to developing our machine learning model, we investigated whether the L3 peak area alone might serve as a more reliable metric for oxidation state classification than the L3/L2 ratio. Figure 6.4 shows that Ni2+ spectra typically exhibit areas between 2.5 - 2.9, while Ni3+ spectra fall between 3.4 - 4.5. The intermediate range of 2.9 - 3.4 corresponds to a mixture of Ni2+ and Ni3+ states. These thresholds were used to annotate degradation regions in the test data. Figure 6.4: Histograms showing the distribution of L3 peak areas for reference sample of Ni2+ (blue) and Ni3+ (orange). 6.2 Challenges in Training Data Preparation for Machine Learning 157 Figure 6.5: Histograms showing the distribution of L3 peak areas for (a) LowStart, (b) LowEnd, and (c) pristine samples. The dashed red lines indicate the boundaries between Ni2+ (2.5-2.9), mixed (2.9-3.4), and Ni3+ (3.4-4.5) states. Insets show the pixel distribution by oxidation state across the sample depth. 158 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification Building on the findings from the previous chapter, the multi-Gaussian fitting in the previous chapter showed that LowStart and LowEnd samples had degradation zones extending ~20 nm and ~10 nm from the surface, respectively, with no observable degradation in the pristine sample. The inset in Figure 6.4a shows that in the LowStart sample, the Ni2+ and mixed valence states are concentrated in the first ~10 nm from the surface, which is shorter than the ~20 nm degradation zone identified through multi-Gaussian analysis. Similarly, the LowEnd sample (Figure 6.5b) shows some Ni2+ and mixed states near the surface, but the extent is less clear. The pristine sample (Figure 6.5c) shows primarily Ni3+ with scattered instances of mixed valence states, consistent with the absence of significant degradation. While this approach showed improved discrimination compared to the L3/L2 ratio method, it still presented limitations. The L3 area metric captures only one aspect of the spectral differences between oxidation states, neglecting the additional contributions from the L2 peak that helps distinguish Ni2+ from Ni3+. However, for the L3/L2 ratio calculations, the integration range must be carefully selected for each spectrum to account for energy shifts between acquisitions, introducing potential inconsistencies. Additionally, the noisier L2 peak increases the risk of integrating background fluctuations rather than real signal. Due to these considerations, the model development henceforth focuses exclusively on the L3 peak for CNN classification. 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification Having established the limitations of traditional analytical approaches and identified the L3 edge as the most informative spectral region, we developed advanced machine learning models to improve oxidation state classification accuracy. This section details the development and evaluation of a convolutional neural network (CNN) approach optimised for EELS spectral analysis. 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification 159 6.3.1 Data Preparation for Machine Learning Before implementing the CNN model, we prepared a comprehensive training dataset from our reference compounds. For each oxidation state (Ni2+ and Ni3+), multiple spectrum images were acquired under identical instrumental conditions as the NMC811 test samples. These spectra underwent consistent preprocessing: 1. Background subtraction using the optimised polynomial fitting approach described in Section 5.3.3 2. Intensity normalisation to unity maximum 3. Cropping to focus specifically on the L3 edge (840 - 867 eV) To enhance the robustness of our machine learning models and account for experimental variations in peak positions, we implemented a data augmentation strategy. The reference spectra were systematically shifted by up to ±7.5eV to simulate variations in energy calibration across different microscopes or acquisition sessions. This ensured that the models would learn to classify based on spectral shape rather than absolute peak position. The ten random spectra for Ni2+ and Ni3+ which included both reference and augmented spectra are shown in Figure 6.6. Figure 6.6: Representative reference spectra for (a) Ni3+ and (b) Ni2+ used in CNN training, including spectral augmentations to simulate energy drift. Each panel shows ten individual spectra (solid lines), with the dotted line indicating the mean spectrum for each oxidation state. Augmentation was applied to improve the model's robustness to experimental variability and ensure generalisability across different acquisition conditions. 160 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification Table 6.1: Summary of CNN training and validation dataset composition. Dataset Component Number of Spectra Ni²⁺ reference spectra (from NiO) 8,000 Ni³⁺ reference spectra (from LiNiO₂) 8,000 Augmented spectra (±7.5 eV energy shifts) 8,000 Total dataset 24,000 Training set (80%) 19,200 Validation set (20%) 4,800 Table 6.1 summarises the composition of the training and validation datasets used for CNN model development. The augmented spectra were generated by systematically shifting the original reference spectra by up to ±7.5 eV to simulate energy calibration variations across different microscope sessions. Stratified sampling ensured balanced representation of both oxidation state classes in the training and validation subsets. 6.3.2 Convolutional Neural Network Architecture To classify the Ni oxidation state from EELS spectra, we developed a one-dimensional convolutional neural network (CNN) that takes the Ni L3-edge spectrum as input and outputs a label as either Ni2+ or Ni3+. As EELS spectrum is a 1D signal, the CNN model uses convolutions along the energy axis to detect important features, such as where a peak begins (onset), how steep it is (slope), and changes in its shape (inflections). Note that performing convolution itself does not extract features, however the CNN model uses convolutional filters to learn the features during training. This is conceptually similar to how CNNs detect textures or edges in images. The goal is for the model to learn the distinct spectral fingerprints of each oxidation state directly from training data. CNNs are particularly effective at identifying patterns in data, such as EELS spectra. As convolutional filters scan across the spectrum, they extract local features like sharp rises, shoulders, or peak drop-offs which are used to inform the model’s classification. By stacking multiple convolutional layers, the network builds up from basic to increasingly complex features, ultimately recognising the characteristic shape and position of the Ni L3 peak. 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification 161 Figure 6.7: Detailed architecture of the CNN model used for nickel oxidation state classification. The network features multiple convolutional layers with varying kernel sizes, an attention mechanism, and a global average pooling branch working in parallel to extract diverse spectral features. Conv1d(x, y, k) represents a 1D convolutional layer with x input channels, y output channels, and k kernel size. FC(x, y) represents a fully connected layer with x input features and y output features. Architecture designed by Kelvin Chan in collaboration with May Ching Lai. 162 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification A visual overview of the CNN model architecture is shown in Figure 6.7. To help readers unfamiliar with machine learning, Table 6.2 summarises the key components of the architecture alongside intuitive analogies. Table 6.2: Glossary of CNN Terms Used in the Architecture276–279 Terminology Technical Function Convolutional Layer 1D (Conv1D) A neural network layer using 1D convolution to extract features from the input data. Input Channel The input to the convolutional layer. The number of input channels is equal to the input’s dimensionality (1D for EELS spectra for the first convolutional layer). Output Channel The output of the convolutional layer. The number of output channels represents the number of features to be learnt by the layer. Filter/Kernel A window used to apply convolution to an input. The values of these filters are weights determined by training the model. Kernel Size The width of the kernel window. It is also the number of data points the filter examines at a time. Dilation Rate The number of data points skipped by the kernel as it slides across the spectrum. Increasing this value allows the kernel to capture broader features. Batch Normalisation Normalises layer inputs to stabilise and speed up learning. ReLU Activation Keeps only positive values to introduce non-linearity, allowing complex features to be learnt. Max-Pooling Reduces dimensionality by keeping strong signals (e.g. spectra features) and removing weak signals (e.g. background noise). Attention Mechanism Highlights important spectral regions by weighting significant features highly. Sigmoid Activation Sigmoid function used to output weights for the attention mechanism Fully Connected Layer A set of neurons that input features and outputs another set of features from analysing the input Dropout Rate Randomly disables nodes during training to reduce overfitting Softmax Activation Converts raw outputs into probability scores for classification. 6.3 Methodology: Machine Learning Approaches for Oxidation State Classification 163 Our CNN architecture shown in Figure 6.7 involves: 1. Main Convolutional Pathway: Three sequential convolutional layers with increasing filter counts (32 → 64 → 128) and decreasing kernel sizes (7 → 5 → 3). Each is followed by batch normalisation, ReLU activation, and max-pooling. This pathway captures hierarchical spectral features ranging from sharp edges to broader trends. 2. Dilated Convolutional Layer: A fourth convolutional layer with 256 filters (kernel size 3, dilation rate 2) expands the receptive field, allowing the model to capture broader spectral context without increasing model complexity. 3. Attention Mechanism: A parallel branch applies a 1D convolution with kernel size of 1, followed by sigmoid activation to generate attention weights. This guides the model to focus on the most relevant regions of the spectrum, which it learns as the model gets trained. 4. Global Average Pooling Branch: This parallel path compresses global spectral information using adaptive pooling and dense layers (256 → 64 nodes) to provide the model with awareness of the entire spectra while the main pathway focuses on the small details. 5. Fully Connected Layers: Outputs from both branches are passed through fully connected layers to analyse the spectral features. Dropout is added to the fully connected layers to reduce overfitting, allowing the model to be more generalised. 6. Output Layer: A final fully connected layer with softmax activation outputs class probabilities for each oxidation state (Ni2+ or Ni3+). This dual-pathway architecture, combining local and global feature extraction with an attention mechanism, was specifically tailored to recognise subtle differences in spectral shape, such as the width, skew, or energy shift of the Ni-L3 peak. These are key indicators of nickel oxidation state, and the CNN’s ability to learn and integrate such multi-scale features makes it particularly suited for this classification task. 6.3.3 Training and Validation Methodology The convolutional neural network (CNN) was trained using the Adam optimisation algorithm, a widely adopted method that updates model parameters efficiently to minimise prediction error.280 A learning rate of 0.001 was selected to balance convergence speed and stability; lower values tend to result in slower but more stable training, while higher values may accelerate convergence at the risk of overshooting optimal solutions. 164 6.4 Classification Results and Discussion To evaluate classification performance, the categorical cross-entropy loss function was used. This metric is particularly suited for multi-class classification problems, comparing predicted probability distributions with true class labels and guiding the model’s learning process.281 To prevent overfitting, where a model performs well on training data but poorly on unseen data, two regularisation strategies were used. Firstly dropout layers were applied to the fully connected portion of the network. These layers randomly deactivated a fraction of neurons during training (with dropout rates of 50% for one layer and 30% for another layer), which encourages the model to generalise rather than memorise.282 Secondly, early stopping was used to automatically halt training once the model's performance on a separate validation set stopped improving, which was after 7 epochs, where an epoch is one complete pass through the training data. The dataset was partitioned into 80% for training and 20% for validation using stratified sampling, ensuring balanced representation of Ni2+ and Ni3+ spectra across both subsets. Training was conducted in batches of 4,096 spectra and capped at 20 epochs, although early stopping terminated the process at 7 epochs due to convergence. As discussed in Section 6.2, attempts to train the model using both L2 and L3 edges led to inconsistent and unreliable classification, with outcomes heavily skewed toward Ni2+ or exhibiting random noise. Improved and stable performance was only achieved when training was restricted to the L3-edge region and spectral shift augmentation was applied. This outcome highlights the critical importance of targeted feature selection and robust preprocessing in the application of machine learning to spectroscopic data. 6.4 Classification Results and Discussion 6.4.1 Accuracy of model The trained CNN model achieved an overall validation accuracy of 98.8%, with perfect classification of Ni2+ spectra and 97.5% accuracy for Ni3+. Figure 6.8 presents the confusion matrix for the validation dataset, showing the distribution of true and predicted labels across 4,800 validation spectra (see Table 6.1 for dataset composition). 6.4 Classification Results and Discussion 165 2+ 3+ Correctly labelled 2+ Incorrectly labelled 2+ T ru e L ab el 2+ 2402 0 2+ (TP) 2+ (FN), 3+ (FP) Incorrectly labelled 3+ Correctly labelled 3+ 3+ 59 2339 2+ (FP), 3+ (FN) 3+ (TP) Predicted Label Figure 6.8: Confusion matrix for CNN classification of nickel oxidation states using EELS spectra. The matrix shows results for 4,800 validation spectra. Diagonal cells represent correct classifications, True Positives (TP): 2,402 Ni2+ spectra and 2,339 Ni3+ spectra were correctly identified. Off-diagonal cells represent misclassifications: 59 Ni3+ spectra were incorrectly labelled as Ni2+, as False Positives (FP) for Ni2+ or False Negatives (FN) for Ni3+, while no Ni2+ spectra were misclassified. Note that reference labels are assumed ground truth and have not been independently verified by complementary techniques such as XAS or XPS. A confusion matrix is a table that summarises how well a classification model performs by comparing its predictions against the actual (true) labels. The matrix is organised with the true labels as rows and the predicted labels as columns. Each cell shows how many samples fell into that combination of true versus predicted class: • True Positives (TP): Spectra correctly identified as their true class (diagonal cells). For Ni2+, TP = 2,402; for Ni3+, TP = 2,339. • False Positives (FP): Spectra incorrectly labelled as a class they do not belong to. For Ni2+ FP = 59 (these were actually Ni3+ spectra misclassified as Ni2+). • False Negatives (FN): Spectra that belonged to a class but were incorrectly assigned elsewhere. For Ni3+, FN = 59 (these Ni3+ spectra were missed and labelled as Ni2+). • True Negatives (TN): In binary classification, this represents samples correctly identified as not belonging to the positive class. 166 6.4 Classification Results and Discussion The confusion matrix reveals that all misclassifications occurred in one direction: 59 spectra (2.5%) of true Ni3+ were incorrectly labelled as Ni2+, while no Ni2+ spectra were misclassified. This asymmetric error pattern suggests a slight bias toward the Ni2+ class, possibly due to: • Edge cases near the decision boundary: Some Ni3+ spectra might exhibit features that partially resemble Ni2+, particularly those with slightly narrower or more symmetric L3 peaks. • Mixed valence contributions: The reference LiNiO2 (nominally Ni3+) may contain small amounts of Ni2+ due to non-stoichiometry or surface reduction, creating ambiguous examples. • Augmentation effects: The data augmentation process, while beneficial for robustness, might have created some spectra that lie very close to the decision boundary. Beyond overall accuracy, additional metrics provide insight into the model's performance for each oxidation state class. These metrics help assess whether the model is equally reliable for both Ni2+ and Ni3+ classification, or whether it favours one class over the other. Table 6.3 presents three complementary metrics calculated from the confusion matrix. To understand these metrics, consider a simple analogy: imagine searching for a specific type of battery defect in a batch of cells. • Precision answers: "Of all the cells I flagged as defective, how many actually were?" High precision means few false alarms. Mathematically, Precision = TP / (TP + FP). • Recall answers: "Of all the truly defective cells, how many did I catch?" High recall means few missed detections. Mathematically, Recall = TP / (TP + FN). • F1-Score is the balanced average of precision and recall, useful when both false alarms and missed detections matter equally. Table 6.3: Classification performance metrics for CNN-based Ni oxidation state determination. Class Precision Recall F1-Score Support Ni2+ 0.976 1.000 0.988 2,402 Ni3+ 1.000 0.975 0.988 2,398 Weighted Average 0.988 0.988 0.988 4,800 6.4 Classification Results and Discussion 167 The classification metrics reveal an asymmetry that has practical implications for degradation mapping. The Ni3+ class achieves perfect precision (1.000): every spectrum the model classified as Ni3+ was genuinely Ni3+, meaning Ni3+ assignments can be trusted completely. The Ni2+ class achieves perfect recall (1.000): the model successfully identified every true Ni2+ spectrum in the validation set, ensuring no degraded regions are missed. The 59 misclassified spectra (Ni3+ incorrectly labelled as Ni2+) affect two metrics: Ni2+ precision drops to 0.976 (meaning 2.4% of pixels labelled as Ni2+ may actually be Ni3+), and Ni3+ recall drops to 0.975 (meaning 2.5% of true Ni3+ spectra were mislabelled). For degradation mapping, this error direction would tend to slightly overestimate Ni2+-like regions rather than miss Ni2+- like spectra, although the true impact will depend on how closely the validation labels reflect oxidation states in the experimental dataset. 6.4.2 Spatial Mapping of Oxidation States in NMC811 Applying the optimised CNN classifier to the NMC811 samples revealed distinct spatial patterns of nickel oxidation states that closely correlate with electrochemical performance differences observed between cycling protocols. Table 6.4 summarises the test dataset characteristics, including the number of spectra classified in each sample. Table 6.4: Test dataset characteristics and CNN classification results. Sample Total Spectra Classified Ni2+ Classified Ni3+ LowStart 874 360 (41.2 %) 514 (58.8 %) LowEnd 1122 103 (9.2 %) 1019 (90.8 %) Pristine 960 62 (6.5 %) 898 (93.5%) Figure 6.9 presents both classification maps (left column) and corresponding probability maps (right column) for the LowStart, LowEnd, and pristine samples. The classification maps show discrete pixel-wise predictions of the dominant oxidation state, with each pixel assigned to either Ni2+ (blue) or Ni3+ (orange) based on the CNN’s output. In contrast, the probability maps 168 6.4 Classification Results and Discussion visualise the model’s prediction confidence for each pixel as a continuous gradient, where blue indicates high confidence for Ni2+, orange indicates high confidence for Ni3+, and white corresponds to ambiguous regions (~50 % confidence), potentially reflecting mixed valence states or transitional zones. These probability maps provide an additional layer of interpretability, allowing assessment of the certainty of predictions across the field of view. Regions near the particle surface, where oxidation states often evolve gradually, frequently exhibit these intermediate probabilities. By visualising both the classification and the associated prediction confidence, we gain a clearer understanding of the spatial redox heterogeneity in cycled NMC811 cathodes. Figure 6.9: CNN-classified oxidation state maps for (a,c,e) discrete classification and (b,d,f) probability visualisation of LowStart, LowEnd, and pristine NMC811 samples, respectively. Blue regions represent Ni2+-rich areas, while orange indicates Ni3+-dominated regions. In the probability maps, the color gradient indicates the model's confidence in classification, with white regions representing approximately equal probability between Ni2+ and Ni3+, suggesting mixed valence states. The x-axis represents distance from the particle surface in nanometers. In the LowStart sample (Figure 6.9a-b), a distinct Ni2+-rich layer extends approximately 20 nm from the surface. The probability map reveals a gradual transition rather than a sharp interface, aligning with the Gaussian-fitting trends in Figure 5.12. This thick, reduced surface zone 6.4 Classification Results and Discussion 169 correlates with the poor electrochemical performance and extensive surface reconstruction observed in HAADF-STEM imaging (Figure 5.15). In contrast, the LowEnd sample (Figure 6.9c-d) displays a thinner ~10 nm surface reduction zone, with limited intermediate regions. This thinner, less extensive degradation layer is consistent with improved electrochemical stability, despite greater mechanical fracturing observed in these samples. The pristine sample (Figure 6.9e-f) shows a near-uniform distribution of Ni3+ with minimal Ni2+ content, confirming the absence of significant surface degradation before cycling. The lack of a Ni2+-rich layer validates the material’s initial stoichiometry and underscores the accuracy of the CNN classification as pristine NMC811 in the discharged state should theoretically contain predominantly Ni3+ with minimal Ni2+.283,284 These spatial patterns align closely with the redox gradients identified using conventional metrics in Chapter 5, such as L3/L2 ratios and oxygen ΔE shifts. However, the CNN offers superior spatial resolution and classification confidence, particularly for LowEnd samples, where traditional methods failed to delineate clear boundaries. This enhanced performance stems from the CNN’s ability to learn from multiple spectral features simultaneously, rather than relying on single-value metrics.141 Across all three samples, the right column (Figure 6.10c, f, i) reveals distinct spectral differences between Ni2+ and Ni3+-classified regions. These differences are most pronounced in the LowStart sample, where the Ni2+-classified spectra (Figure 6.10a) closely resemble the sharp, symmetric L3 peak of the reference Ni2+ spectrum, indicating near-complete reduction at the particle surface. In the LowEnd and pristine samples (Figure 6.10d and g), the Ni2+-classified spectra exhibit broader peaks than the reference, suggesting that these regions represent mixed-valence environments rather than fully reduced Ni2+. This is particularly evident in the LowEnd sample, where the average profile (Figure 6.10f) shows less pronounced peak shifts compared to LowStart. 170 6.4 Classification Results and Discussion Figure 6.10: Comparative spectral analysis of Ni oxidation states in NMC811 samples. Left column (a,d,g): Average spectra of pixels classified as Ni2+ (blue) compared to Ni2+ reference (red). Middle column (b,e,h): Average spectra of pixels classified as Ni3+ (orange) compared to Ni3+ reference (red). Right column (c,f,i): Comparison of mean spectra for all pixels classified as Ni2+ (blue) versus Ni3+ (orange) within each sample. Top row: LowStart sample; Middle row: LowEnd sample; Bottom row: Pristine sample. All test spectra were energy-aligned with reference spectra to facilitate direct comparison of peak shapes. 6.4 Classification Results and Discussion 171 Figure 6.10 compares averaged spectra from pixels classified as Ni2+ or Ni3+ against reference spectra. Each solid coloured line represents the mean spectrum, calculated by averaging all individual spectra within that classification category. For example, the solid blue line in Figure 6.10a is the average of 360 individual EELS spectra that the CNN classified as Ni2+ in the LowStart sample (see Table 6.4 for counts in each sample). The shaded regions (halos) surrounding each mean line visualise the spread of individual spectra around that average. At each energy channel, the shaded band shows ± one standard error, calculated as: 𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝐸𝑟𝑟𝑜𝑟 = 𝜎 √𝑛 Equation 23 where 𝜎 is the standard deviation of intensity at that energy channel across all 𝑛 spectra in the category. The width of the shaded bands reflects the spectral diversity within each classification category. Narrow bands indicate that the CNN grouped together spectra with highly consistent shapes, while broader bands indicate greater variation among the classified spectra. The narrow shaded bands for Ni3+-classified pixels in the pristine and LowEnd samples indicate that these spectra are highly consistent with each other, reflecting the chemical uniformity of bulk layered NMC811. In contrast, the broader bands for Ni2+-classified pixels in the LowStart sample suggest that the CNN is classifying a more heterogeneous set of spectra, likely spanning a range from fully reduced Ni2+ to partially reduced transitional states that still exceed the classification threshold. This is consistent with the expected chemical gradient within surface reconstruction layers, where reduction proceeds progressively rather than as a sharp boundary. This gradual variation in spectral features reflects the real-world behaviour of surface degradation in Ni-rich layered oxides, where the structural transformation from the layered phase (dominated by Ni3+) to the rock-salt phase (enriched in Ni2+) does not proceed through a sharp, two-phase boundary. Instead, it progresses through a disordered spinel-like intermediate phase, often characterised by partial reduction of nickel and oxygen loss.283,285,286 The CNN captures this subtle evolution through its probability outputs, which reveal regions with ambiguous classification that likely correspond to these transitional environments. This makes the CNN a more chemically realistic and sensitive tool compared to traditional binary methods, which may oversimplify such complex degradation pathways. These redox gradients are strongly linked to electrochemical performance. The thicker, more fully reduced Ni2+-rich surface layer in LowStart samples likely increases interfacial impedance and limits lithium-ion diffusion, contributing to rapid capacity fade. In contrast, 172 6.4 Classification Results and Discussion the thinner degradation layer in LowEnd samples corresponds with improved electrochemical stability, despite the presence of structural cracks. This supports recent findings by Lee et al.287 that surface chemical degradation, rather than mechanical fracturing, plays a more critical role in long-term cell degradation. Thus, the CNN-based classification framework helps to establish a direct mechanistic link between nanoscale redox processes and macroscopic battery performance. By providing high- resolution maps of oxidation state changes, it offers practical insight for designing more resilient cycling protocols, suggesting that strategies minimising high-voltage exposure (as in the LowEnd protocol) may be more effective in maintaining cathode integrity than approaches focusing solely on mechanical stress mitigation. 6.4.3 Limitations and Future Opportunities for CNN-Based Classification While the CNN model demonstrates strong performance in resolving spatial oxidation state distributions and capturing subtle redox transitions in NMC811, several areas offer opportunities for further development and refinement. One current constraint lies in the binary classification scheme (Ni2+ vs Ni3+), which does not explicitly account for mixed or intermediate valence states, features that are frequently encountered in real battery materials. In practical systems, these intermediate states may arise from incomplete redox reactions, surface reconstruction, or prolonged cycling, and can have markedly different implications depending on whether they are reversible/metastable (e.g., transient Ni2+/Ni3+ mixtures) or irreversible/stable (e.g., persistent Ni2+ formation due to surface degradation). The probability maps generated by the CNN already offer some insight into these transitional regions, but more continuous modelling approaches could further improve this. To enable this, it would be valuable to obtain reference spectra from materials with well-characterised mixed valence states or intermediate oxidation levels (e.g., Ni2.5+, Ni3.5+), as well as spectra from Ni4+-rich environments. Alternatively, synthetic datasets could be generated by mixing experimentally acquired spectra in controlled proportions or by simulating EELS responses based on theoretical models. Training the CNN on such extended datasets could allow for accurate classification across the full Ni2+ to Ni4+ range, including spatially heterogeneous or partially reduced regions commonly found in aged electrodes. In addition to valence ambiguity, plural scattering presents another key challenge. As sample thickness increases, inelastic scattering events accumulate, broadening the energy-loss peaks and distorting their shapes. For example, in Mn3O₄, increasing the thickness from t/λ = 0.3 to 6.5 Relating EELS-CNN Findings to Other Spectroscopic Techniques 173 1.0 causes the L3 peak ratio to shift from 2.65 to 2.85 due to broadening and plural contributions.130,288 While deconvolution techniques like Fourier-log deconvolution can correct for plural scattering, they often reduce the signal-to-noise ratio, especially in low-dose or beam-sensitive systems.289,290 A promising alternative would be to train the CNN using synthetic spectra generated at different 𝑡/𝜆 conditions. This would allow the model to learn the underlying spectral signatures of each oxidation state while remaining robust to thickness- related distortions, a strategy already demonstrated successfully by MnEdgeNet141 for manganese L-edges. Incorporating similar modelling for Ni could eliminate the need for deconvolution and streamline analysis. The model’s overall performance also depends on the quality and consistency of the training data. Variability in background subtraction, energy alignment, and reference material purity can all affect classification accuracy. Incorporating standardised preprocessing pipelines or tailored data augmentation techniques could improve robustness and generalisability across instruments and datasets. Addressing these limitations will be key to expanding the CNN-based classification framework to more complex transition metal chemistries, enabling multielement redox mapping, and supporting real-time or high-throughput applications in operando battery research. 6.5 Relating EELS-CNN Findings to Other Spectroscopic Techniques 6.5.1 Similarities Between EELS and XAS for Oxidation State Analysis Electron Energy Loss Spectroscopy (EELS) and X-ray Absorption Spectroscopy (XAS) both probe transition metal L-edge transitions, corresponding to 2p → 3d excitations.291–293 Despite differing excitation mechanisms - inelastic scattering for EELS and photon absorption for XAS, the resulting spectral features are closely aligned. In both cases, oxidation states influence L2,3 peak shape, position, and intensity. This spectral similarity forms the basis for extending deep learning models developed for EELS to XAS analysis. EELS offers nanometre-scale spatial resolution ideal for mapping localised redox gradients. On the other hand, XAS provides depth-tunable analysis, with TEY (Total Electron Yield) probing near-surface regions (~5-10 nm) and FY (Fluorescence Yield) reaching deeper into the material (~100-200 nm).291 Together, these techniques capture complementary dimensions of degradation in layered cathodes like NMC811. 174 6.5 Relating EELS-CNN Findings to Other Spectroscopic Techniques 6.5.2 Adapting CNN Classification Across EELS and XAS: Complementarity and Compatibility The CNN architecture developed for EELS spectra is well-suited in principle to XAS data, given the structural similarities between their respective L-edge spectra. Both techniques yield one-dimensional spectral profiles shaped by the same 2p → 3d electronic transitions, and both benefit from common preprocessing steps such as background subtraction, normalisation, and energy alignment. Spectral augmentation during training improves robustness to energy misalignment; however, cross-technique transfer would still require careful re-calibration of preprocessing and likely fine-tuning on XAS spectra to mitigate domain shift (e.g., resolution, background and detection-mode effects). Although EELS and XAS differ in excitation mechanism and scale, the spectral features that distinguish oxidation states, such as peak shape, width, and energy position, are highly conserved. As such, the convolutional layers, attention mechanisms, and feature extraction strategies of the CNN remain effective when applied to either technique, with minimal architectural modification. Beyond spectral similarity, EELS and XAS offer complementary practical advantages. EELS excels at nanoscale mapping of redox heterogeneity in individual particles, while XAS provides improved signal stability, is non-destructive, and is ideally suited for operando studies and thicker samples. When used together, they provide a more comprehensive picture of degradation, capturing both lateral and depth-dependent redox behaviour. Taken together, these strengths position CNNs as flexible tools for cross-technique oxidation state analysis. A model trained on either EELS or XAS data, and ideally both, could support unified, high-confidence classification across different spatial scales and experimental platforms. 6.5.3 Prospects for Unifying EELS and XAS Through CNN Classification A recent study by the Weatherup group262 demonstrated the compatibility of EELS and XAS for probing redox evolution in LiNiO2 (LNO) following high-voltage cycling. EELS revealed a ~200 nm Ni2+-rich surface layer transitioning to Ni4+ in the bulk. Complementary XAS measurements confirmed this gradient: TEY-XAS detected strong surface reduction, while FY- XAS revealed both Ni2+ features and molecular O2 in the near-surface region. Although CNNs were not employed in that study, the agreement between techniques reinforces the feasibility of applying a shared CNN framework across both methods, building on the architectural compatibility discussed earlier. 6.6 Conclusion 175 Extending CNN-based classification to multi-technique datasets presents an exciting opportunity to unify redox mapping across spatial and depth dimensions. A single CNN model trained on both EELS and XAS reference spectra for Ni2+, Ni3+, and Ni4+ could enable standardised, technique-independent oxidation state classification, eliminating inconsistencies from manual peak fitting and reducing analyst bias. The CNN developed here is already robust to variation in noise, background levels, and energy drift, making it well- suited for transfer to XAS datasets without major architectural changes. When applied to EELS line scans or operando XAS experiments, a unified CNN could generate oxidation state maps that combine both lateral and depth-resolved information. For NMC811, this would enable clearer correlations between surface degradation and bulk redox evolution, shedding light on degradation pathways under different cycling protocols. In addition, CNNs offer the potential for high-throughput analysis, capable of detecting fine- scale chemical variations such as gradual redox shifts or partial reduction zones, that may be overlooked in manual workflows. Looking forward, CNN models could be expanded to jointly analyse multiple absorption edges, such as combining Ni L-edge and O K-edge data, to capture the evolution of metal- oxygen hybridisation. This is particularly relevant for understanding capacity fade and voltage hysteresis in layered cathode materials. Such integrated models would support predictive analysis of cathode performance and inform the rational design of more durable, high-energy battery chemistries. 6.6 Conclusion Initial analysis using the conventional L3/L2 ratio revealed general redox trends but was hindered by spectral overlap and sensitivity to noise. Using the L3 peak area improved classification performance and was less affected by noise; however, this method relied on a single integrated value, making it insensitive to more subtle differences in spectral shape. As a result, it could not capture finer distinctions between oxidation states, especially in mixed or transitional regions. To overcome these limitations, we developed a CNN model trained solely on the Ni L3-edge region of EELS spectra, using energy-shift augmentation to enhance generalisability. The resulting model achieved a validation accuracy of 98.8% and enabled pixel-level classification of Ni2+ and Ni3+ states with high spatial precision. Table 6.5 summarises the relative performance of each method explored in this work. Together, these findings establish CNN-enhanced EELS as a powerful platform for analysing nanoscale redox transformations in complex battery materials. The ability to perform pixel- level classification enables more precise visualisation of oxidation gradients and degradation 176 6.6 Conclusion fronts, delivering insights essential for understanding and improving cathode behaviour under different cycling protocols. Table 6.5: Comparison of different approaches for nickel oxidation state classification using EELS. Method Accuracy Pros Cons L3/L2 Ratio Low Simple to compute Overlap of states, noise sensitive L3 Area Moderate Less noisy than ratio Still ambiguous; ignores spectral shape CNN (L3-only) High (98.8%) Robust, generalisable Requires labelled data; black-box architecture Future directions for this research include: 1. Simultaneous classification across elements: Expanding the CNN framework to simultaneously classify oxidation states of Ni, Mn, and Co would provide a more holistic view of redox interactions in multi-cation cathode systems like NMC811. 2. Operando and time-resolved applications: Applying this approach to operando EELS or XAS datasets would allow real-time tracking of redox evolution during cycling, uncovering transient states and enabling predictive modelling of degradation pathways. 3. Unified models across modalities: Combining EELS and XAS input into a single CNN framework would support consistent cross-technique classification, enhancing interpretability and data integration in multi-scale, multi-modal studies. 4. Integration with other data streams: Coupling CNN-classified oxidation state maps with electrochemical data, mechanical strain mapping, or ion diffusion models could provide comprehensive structure-property-function correlations. By bridging atomic-resolution spectroscopy with machine learning and electrochemical understanding, this work represents a step toward intelligent diagnostics in battery research. It not only advances our mechanistic understanding of degradation but also offers practical guidance for optimising electrode design and charging strategies in next-generation lithium- ion batteries. 7.1 Summary of Methodological Contributions 177 Chapter 7 Conclusions and Future Works This thesis has developed and validated a suite of analytical methodologies for characterising degradation mechanisms and redox evolution in NMC811 cathode materials. By integrating advanced electron microscopy, spectroscopic analysis, and data-driven modelling across multiple length scales, the work establishes reproducible frameworks for extracting quantitative information from complex battery materials. Overall, the results show that degradation in Ni-rich layered cathodes is strongly protocol-dependent and manifests as coupled changes in near-surface redox chemistry and transport-relevant microstructure. The methodological contributions span three-dimensional microstructural characterisation, spatially-resolved oxidation state mapping, and machine learning-assisted spectral classification. This chapter summarises the key methodological advances, discusses their broader applicability, and identifies directions for future development. 7.1 Summary of Methodological Contributions 7.1.1 Three-Dimensional Microstructural Characterisation The first methodological contribution of this thesis is the development of a FIB-SEM tomography workflow integrated with supervised machine learning segmentation. Implemented using Dragonfly software, this approach enables more consistent segmentation of electrode microstructures at sufficient resolution to distinguish carbon-binder domains from active material and pore phases. This capability addresses a common challenge in electrode characterisation, where conventional thresholding methods struggle to separate phases with similar greyscale intensities. The workflow was validated through systematic comparison of three independent tortuosity calculation methods: graph-based path analysis (Dragonfly), finite-difference simulation (TauFactor), and electrochemical impedance spectroscopy (EIS). This multi-method approach provides cross-validation of microstructural parameters and enables quantification of measurement uncertainty. The methodology was applied to investigate single-crystal (SC) and polycrystalline (PC) NMC811 electrodes across a range of calendering conditions, revealing distinct structural responses to mechanical processing. 178 7.1 Summary of Methodological Contributions Key findings from the microstructural analysis include calendering targets suggested by the conditions studied here: ~25% porosity for SC electrodes, which retain structural integrity while developing preferential transport pathways, and ~ 35% porosity for PC electrodes, which exhibit increasing intergranular cracking at higher compaction levels. The quantification of tortuosity anisotropy demonstrated that directional transport effects, often overlooked in conventional characterisation, can influence electrode performance. These results were supported by correlations with electrochemical testing, establishing structure- performance relationships that can inform electrode manufacturing optimisation. 7.1.2 EELS Analysis and Multi-Gaussian Fitting The second methodological contribution is the development of a systematic EELS acquisition and analysis protocol for probing nanoscale oxidation state variations in NMC811. This work addresses practical challenges associated with acquiring quantitative EELS data from mechanically fragile cathode materials. Through theoretical analysis and experimental validation, a practical thickness criterion (𝑡/𝜆 ≈ 0.5) was adopted to balance signal quality against plural scattering artefacts. A modified FIB lamella mounting approach was developed to enable reliable preparation of ultrathin specimens (40-65 nm) suitable for high-quality EELS acquisition. A multi-Gaussian fitting approach was designed to analyse overlapping spectral features in the transition metal L2,3 edges. This approach is particularly relevant for EELS datasets acquired using non-monochromated systems with medium energy dispersion (0.3 eV/channel), where multiplet splitting cannot be directly resolved. Given that non- monochromated microscopes are more commonly available in routine characterisation facilities, the fitting protocol enhances the accessibility of quantitative oxidation state analysis. The pipeline was benchmarked against known reference standards and made robust to common experimental issues including energy drift, plural scattering, and limited energy resolution, enabling consistent extraction of L3/L2 ratios and ΔE peak separations. Application of this methodology to cycled NMC811 electrodes revealed protocol-dependent differences in surface degradation. The LowStart charging protocol, which prolongs the constant voltage phase at high voltage, produced a surface layer (~20 nm) dominated by Ni3+ to Ni2+ reduction, consistent with surface reconstruction. The LowEnd protocol, which gradually tapers current at end of charge, preserved a more intact layered structure with a thinner chemically altered region (~10 nm). These observations were corroborated by high- resolution STEM imaging showing progressive phase transitions from layered to spinel and rock-salt structures in degraded surface regions. 7.2 Recommendations for Microstructural Characterisation 179 7.1.3 Machine Learning Classification of Oxidation States The third methodological contribution is the development of a convolutional neural network (CNN) architecture for automated classification of nickel oxidation states from EELS spectra. Trained on the Ni L3-edge region, the model achieved 98.8% validation accuracy in distinguishing Ni2+ and Ni3+ states. This approach enables high-throughput, pixel-by-pixel mapping of oxidation states at nanometre resolution, addressing limitations of manual spectral analysis which can be subjective and time-consuming for large datasets. The CNN generates probability maps that visualise prediction confidence as a continuous gradient, capturing the gradual nature of redox transitions at particle surfaces. This provides an additional layer of interpretability beyond discrete classification, allowing identification of transition zones where mixed oxidation states may be present. The model was validated against independent test spectra and shown to produce consistent results across different particle regions and cycling conditions. Together, the Gaussian fitting and CNN-based approaches provide complementary tools for EELS analysis: the former offers transparent, physically-interpretable fitting parameters suitable for detailed mechanistic studies, while the latter enables rapid screening of large spectral datasets. Both methods contribute to making quantitative oxidation state mapping more accessible and reproducible. 7.2 Recommendations for Microstructural Characterisation Several directions would strengthen the microstructural characterisation framework developed in this thesis. Access to plasma FIB capabilities would enable acquisition of significantly larger tomographic volumes, providing more statistically robust sampling of heterogeneous electrode structures. The current serial-sectioning FIB-SEM workflow, while providing sufficient resolution to resolve carbon-binder domains, is limited to representative volumes of approximately 15-20 𝜇𝑚 in each dimension. Plasma FIB could extend this to volumes of 50-100 𝜇𝑚 , capturing rare microstructural features and improving statistical confidence in derived parameters. Comparative studies between X-ray computed tomography (CT) and FIB-SEM would provide valuable insights into the trade-offs between field of view and spatial resolution. Key questions include the transferability of trained segmentation models between imaging modalities, and the impact of resolving carbon-binder domains on calculated tortuosity values. Such cross-validation would help establish guidelines for selecting appropriate characterisation techniques based on the specific questions being addressed. 180 7.3 Recommendations for Spectroscopic Analysis The tortuosity anisotropy findings suggest opportunities for deliberately engineering directional transport pathways. Future work could investigate controlled particle alignment techniques during electrode fabrication, such as magnetic field-assisted processing or gradient porosity structures. The calendering-induced preferential orientation observed in SC electrodes indicates that transport anisotropy can be manipulated through manufacturing parameters, potentially offering a route to performance optimisation. A limitation of the current study is the use of single electrodes for some conditions due to equipment and material availability. Future studies should incorporate replicate measurements across multiple electrodes fabricated under nominally identical conditions to properly quantify batch-to-batch variability and establish confidence intervals for reported parameters. 7.3 Recommendations for Spectroscopic Analysis The Gaussian fitting methodology offers an accessible entry point for researchers less familiar with spectral analysis, with scope for further development. Interactive implementation through graphical notebooks with drag-to-fit interfaces would lower the barrier to adoption. Expansion of reference datasets to include Mn and Co L-edges, as well as intermediate nickel oxidation states (e.g., Ni2.5+), would strengthen the quantitative basis for redox analysis in complex multi-cation cathode systems. For the CNN classification approach, priorities for future development include improving robustness to variations in specimen thickness, plural scattering intensity, energy calibration, and noise levels. These factors vary between microscope sessions and specimens, and a model trained on data from one set of conditions may not generalise well to others. Transfer learning approaches and data augmentation strategies could help address this challenge. Extension of the CNN framework to simultaneous multi-element classification (Ni, Mn, Co) would provide a more complete picture of cationic redox interactions in NMC materials. This is particularly relevant for understanding charge compensation mechanisms and identifying conditions under which specific transition metals become electrochemically active or undergo reduction. Integration of EELS and X-ray absorption spectroscopy (XAS) analysis within a unified machine learning framework would enable consistent oxidation state classification across spatial scales, bridging local (EELS, nanometer-ranged) and bulk-averaged XAS sensitivities (depending on mode, from near-surface TEY to deeper 𝜇𝑚 -ranged FY modes). This multimodal approach could improve confidence in interpretation and help identify when local observations are representative of bulk behaviour. 7.4 Open-Source Software and Reproducibility 181 Coupling oxidation state maps with complementary data streams, such as electrochemical profiles, strain mapping from geometric phase analysis, or ion diffusion coefficients from tracer experiments, would support deeper structure-property correlation and provide more mechanistic insight into degradation processes. 7.4 Open-Source Software and Reproducibility As this thesis heavily relies on Python-based data analysis, it is important to underscore the value of open-source software in advancing modern scientific research. While research publications are often meticulous in detailing sample preparation and data acquisition, they frequently fall short in providing transparent and reproducible descriptions of data processing and analysis workflows. Many in the microscopy community come from experimental backgrounds and may lack formal coding expertise. When new analytical pipelines are developed, particularly those addressing specific challenges in quantification or image processing, it is therefore crucial that they are shared openly. While a growing number of open-source tools are now available, many remain underutilised due to the lack of clear documentation and practical example notebooks. Without these supporting resources, even well-designed code can remain inaccessible to much of the community. Publishing code via platforms such as GitHub, along with user-friendly documentation and demonstrative notebooks, facilitates broader adoption and empowers users of varying technical proficiency. Such practices not only improve reproducibility but also promote collaborative growth within the research community. This shift is particularly important in interdisciplinary fields like battery research, where it is uncommon for a single research group to possess full expertise across sample fabrication, data acquisition, and advanced data analysis. Encouraging open sharing of tools and workflows, paired with a commitment to usability, supports a more integrated scientific ecosystem, one in which collaboration, transparency, and reproducibility drive progress. 7.5 Final Reflection This thesis has been a journey of combining imaging, spectroscopy, and machine learning to better understand degradation in NMC811 cathodes. What stood out most throughout was the importance of collaboration. I started my PhD with no experience in electron microscopy, and having technicians and postdocs willing to invest their time to train me across multiple instruments was invaluable in helping me become a self-sufficient microscopist. 182 7.5 Final Reflection I also began this project with no background in coding, and each chapter became a learning curve that gradually built my confidence and skills. Collaborators who were more experienced in data analysis, and generous with their time and guidance, played a key role in that development. Without direct access to electrode fabrication or deep electrochemical expertise, I relied heavily on support from others. The Faraday Institution provided not only the data and resources, but also a platform to connect with researchers across disciplines, enabling the collaborative environment that made this work possible. In the end, science is a collective effort. I hope the insights developed here, from methods to quantify transport networks and redox gradients, to tools for EELS analysis and classification, can support others in their own research and contribute to continued progress in battery science. References 183 References 1. Global EV Outlook 2024 – Analysis - IEA. https://www.iea.org/reports/global-ev- outlook-2024. 2. Ding, Y., Cano, Z. P., Yu, A., Lu, J. & Chen, Z. Automotive Li-Ion Batteries: Current Status and Future Perspectives. Electrochemical Energy Reviews 2019 2:1 2, 1–28 (2019). 3. Electric car sales, 2012-2024 – Charts – Data & Statistics - IEA. https://www.iea.org/data- and-statistics/charts/electric-car-sales-2012-2024. 4. Battlebury, D. R. A high performance lead–acid battery for EV applications. J Power Sources 80, 7–11 (1999). 5. General Motors’ EV1. https://www.thehenryford.org/explore/blog/general-motors-ev1. 6. Goodenough, J. B., Mizushima, K. & Takeda, T. Solid-solution oxides for storage-battery electrodes. Jpn J Appl Phys 19, 305–313 (1980). 7. Savina, A. A. & Abakumov, A. M. Benchmarking the electrochemical parameters of the LiNi0.8Mn0.1Co0.1O2 positive electrode material for Li-ion batteries. Heliyon 9, e21881 (2023). 8. Polestar 3 | Polestar UK. https://www.polestar.com/uk/polestar-3/ 9. Air Grand Touring | Lucid Motors. https://lucidmotors.com/air-grand-touring. 10. Model 3 | Tesla United Kingdom. https://www.tesla.com/en_gb/model3. 11. Electric vehicle battery prices are expected to fall almost 50% by 2026 | Goldman Sachs. https://www.goldmansachs.com/insights/articles/electric-vehicle-battery-prices-are- expected-to-fall-almost-50-percent-by-2025. 12. Trends in electric vehicle batteries – Global EV Outlook 2024. https://www.iea.org/reports/global-ev-outlook-2024/trends-in-electric-vehicle- batteries (2024). 13. Sun, Y. K., Lee, D. J., Lee, Y. J., Chen, Z. & Myung, S. T. Cobalt-free nickel rich layered oxide cathodes for lithium-ion batteries. ACS Appl Mater Interfaces 5, 11434–11440 (2013). 14. 2025 BMW iX3 - Mid-size Family Electric SUV: Price, Range. https://www.bmw.co.uk/en/all-models/x-models/iX3/2021/bmw-ix3-highlights.html 184 References 15. Schreiber, M. et al. Understanding lithium-ion battery degradation in vehicle applications: Insights from realistic and accelerated aging tests using Volkswagen ID.3 pouch cells. J Energy Storage 112, 115357 (2025). 16. Zhang, S. S. Understanding of performance degradation of LiNi0.80Co0.10Mn0.10O2 cathode material operating at high potentials. Journal of Energy Chemistry 41, 135–141 (2020). 17. Dose, W. M. et al. Onset Potential for Electrolyte Oxidation and Ni-Rich Cathode Degradation in Lithium-Ion Batteries. ACS Energy Lett 7, 3524–3530 (2022). 18. Björklund, E. et al. Cycle-Induced Interfacial Degradation and Transition-Metal Cross- Over in LiNi0.8Mn0.1Co0.1O2-Graphite Cells. Chemistry of Materials 34, 2034–2048 (2022). 19. Dose, W. M. et al. Electrolyte Reactivity at the Charged Ni-Rich Cathode Interface and Degradation in Li-Ion Batteries. ACS Appl Mater Interfaces 14, 13206–13222 (2022). 20. Hausbrand, R. et al. Fundamental degradation mechanisms of layered oxide Li-ion battery cathode materials: Methodology, insights and novel approaches. Materials Science and Engineering: B 192, 3–25 (2015). 21. Kim, J. H., Ryu, H. H., Kim, S. J., Yoon, C. S. & Sun, Y. K. Degradation Mechanism of Highly Ni-Rich Li[NixCoyMn1- x- y]O2 Cathodes with x > 0.9. ACS Appl Mater Interfaces 11, 30936–30942 (2019). 22. Jiang, M., Danilov, D. L., Eichel, R. A. & Notten, P. H. L. A Review of Degradation Mechanisms and Recent Achievements for Ni-Rich Cathode-Based Li-Ion Batteries. Adv Energy Mater 11, 2103005 (2021). 23. Sun, H. H. & Manthiram, A. Impact of microcrack generation and surface degradation on a nickel-rich layered Li[Ni0.9Co0.05Mn0.05]O2 Cathode for Lithium-Ion Batteries. Chemistry of Materials 29, 8486–8493 (2017). 24. Dixit, M., Markovsky, B., Schipper, F., Aurbach, D. & Major, D. T. Origin of Structural Degradation during Cycling and Low Thermal Stability of Ni-Rich Layered Transition Metal-Based Electrode Materials. Journal of Physical Chemistry C 121, 22628–22636 (2017). 25. Liu, Z., Chen-Wiegart, Y. C. K., Wang, J., Barnett, S. A. & Faber, K. T. Three-Phase 3D Reconstruction of a LiCoO2 Cathode via FIB-SEM Tomography. Microscopy and Microanalysis 22, 140–148 (2016). 26. Hwang, S. et al. Determination of the mechanism and extent of surface degradation in Ni-based cathode materials after repeated electrochemical cycling. APL Mater 4, (2016). References 185 27. Li, X. et al. Unravelling the Chemistry and Microstructure Evolution of a Cathodic Interface in Sulfide-Based All-Solid-State Li-Ion Batteries. ACS Energy Lett 4, 2480–2488 (2019). 28. Nitta, N., Wu, F., Lee, J. T. & Yushin, G. Li-ion battery materials: Present and future. Materials Today 18, 252–264 (2015). 29. Morzy, J. K. Electron microscopy study of degradation mechanisms in Ni-rich transition metal oxide cathodes for Li-ion batteries. https://doi.org/10.17863/CAM.94386 (2023) doi:10.17863/CAM.94386. 30. Liu, C., Neale, Z. G. & Cao, G. Understanding electrochemical potentials of cathode materials in rechargeable batteries. Materials Today 19, 109–123 (2016). 31. Khan, Z. A., Agnaou, M., Sadeghi, M. A., Elkamel, A. & Gostick, J. T. Pore Network Modelling of Galvanostatic Discharge Behaviour of Lithium-Ion Battery Cathodes. J Electrochem Soc 168, 070534 (2021). 32. Zacharias, N. A. et al. Direct Measurements of Effective Ionic Transport in Porous Li- Ion Electrodes. J Electrochem Soc 160, A306–A311 (2013). 33. Matsuzaki, K., Shikazono, N. & Kasagi, N. Three-dimensional numerical analysis of mixed ionic and electronic conducting cathode reconstructed by focused ion beam scanning electron microscope. J Power Sources 196, 3073–3082 (2011). 34. DuBeshter, T., Sinha, P. K., Sakars, A., Fly, G. W. & Jorne, J. Measurement of Tortuosity and Porosity of Porous Battery Electrodes. J Electrochem Soc 161, A599–A605 (2014). 35. Thorat, I. V. et al. Quantifying tortuosity in porous Li-ion battery materials. J Power Sources 188, 592–600 (2009). 36. Li, W., Erickson, E. M. & Manthiram, A. High-nickel layered oxide cathodes for lithium- based automotive batteries. Nature Energy 2020 5:1 5, 26–34 (2020). 37. Ciez, R. E. & Whitacre, J. F. Examining different recycling processes for lithium-ion batteries. Nat Sustain 2, 148–156 (2019). 38. Thomas, M. G. S. R., David, W. I. F., Goodenough, J. B. & Groves, P. Synthesis and structural characterization of the normal spinel Li[Ni2]O4. Mater Res Bull 20, 1137–1146 (1985). 39. Ngala, J. K. et al. The synthesis, characterization and electrochemical behavior of the layered LiNi0.4Mn0.4Co0.2O2 compound. J Mater Chem 14, 214–220 (2004). 186 References 40. Yang, S. et al. High Tap Density Spherical Li[Ni0.5Mn0.3Co0.2]O2 Cathode Material Synthesized via Continuous Hydroxide Coprecipitation Method for Advanced Lithium-Ion Batteries . International Journal of Electrochemistry 2012, 1–9 (2012). 41. Cao, H., Zhang, Y., Zhang, J. & Xia, B. Synthesis and electrochemical characteristics of layered LiNi0.6Co0.2Mn0.2O2 cathode material for lithium ion batteries. Solid State Ion 176, 1207–1211 (2005). 42. Rossen, E., Jones, C. D. W. & Dahn, J. R. Structure and electrochemistry of LixMnyNi1−yO2. Solid State Ion 57, 311–318 (1992). 43. Chang, C. C., Scarr, N. & Kumta, R. N. Synthesis and electrochemical characterization of LiMO2 (M=Ni, Ni0.75Co0.25) for rechargeable lithium ion batteries. Solid State Ion 112, 329–344 (1998). 44. Freitas, B. G. A., Siqueira, J. M., da Costa, L. M., Ferreira, G. B. & Resende, J. A. L. C. Synthesis and characterization of LiCoO2 from different precursors by Sol-Gel method. J Braz Chem Soc 28, 2254–2266 (2017). 45. Wu, Z. et al. Ni-rich cathode materials for stable high-energy lithium-ion batteries. Nano Energy 126, 109620 (2024). 46. Li, M. & Lu, J. Cobalt in lithium-ion batteries. Science (1979) 367, 979–980 (2020). 47. Xu, B., Qian, D., Wang, Z. & Meng, Y. S. Recent progress in cathode materials research for advanced lithium ion batteries. Materials Science and Engineering: R: Reports 73, 51– 65 (2012). 48. Bianchini, M., Roca-Ayats, M., Hartmann, P., Brezesinski, T. & Janek, J. There and Back Again—The Journey of LiNiO2 as a Cathode Active Material. Angewandte Chemie International Edition 58, 10434–10458 (2019). 49. Mu, L. et al. Propagation topography of redox phase transformations in heterogeneous layered oxide cathode materials. Nature Communications 2018 9:1 9, 1–8 (2018). 50. Myung, S. T. et al. Nickel-Rich Layered Cathode Materials for Automotive Lithium-Ion Batteries: Achievements and Perspectives. ACS Energy Lett 2, 196–223 (2017). 51. Liu, Z., Yu, A. & Lee, J. Y. Synthesis and characterization of LiNi1−x−yCoxMnyO2 as the cathode materials of secondary lithium batteries. J Power Sources 81–82, 416–419 (1999). References 187 52. Cho, J., Kim, T.-J., Kim, J., Noh, M. & Park, B. Synthesis, thermal, and electrochemical properties of AIPO 4-coated LiNi0.8Co0.1Mn0.1O 2 cathode materials for a Li-ion cell. J Electrochem Soc 151, A1899–A1904 (2004). 53. Märker, K., Reeves, P. J., Xu, C., Griffith, K. J. & Grey, C. P. Evolution of Structure and Lithium Dynamics in LiNi0.8Mn0.1Co0.1O2 (NMC811) Cathodes during Electrochemical Cycling. Chemistry of Materials 31, 2545–2554 (2019). 54. Zheng, J., Kan, W. H. & Manthiram, A. Role of Mn content on the electrochemical properties of Nickel-Rich layered LiNi0.8-xCo0.1Mn0.1+xO2 (0.0 ≤ x ≤ 0.08) cathodes for lithium-ion batteries. ACS Appl Mater Interfaces 7, 6926–6934 (2015). 55. Kim, H. G., Myung, S. T., Lee, J. K. & Sun, Y. K. Effects of manganese and cobalt on the electrochemical and thermal properties of layered Li[Ni0.52Co0.16+xMn0.32−x]O2 cathode materials. J Power Sources 196, 6710–6715 (2011). 56. Xiao, Y. et al. Insight into the origin of lithium/nickel ions exchange in layered Li(NixMnyCoz)O2 cathode materials. Nano Energy 49, 77–85 (2018). 57. Nam, K. W. et al. Combining In Situ Synchrotron X-Ray Diffraction and Absorption Techniques with Transmission Electron Microscopy to Study the Origin of Thermal Instability in Overcharged Cathode Materials for Lithium-Ion Batteries. Adv Funct Mater 23, 1047–1063 (2013). 58. Ryu, H. H., Park, K. J., Yoon, C. S. & Sun, Y. K. Capacity fading of ni-rich li[NixCoyMn1- x-y]O2 (0.6 ≤ x ≤ 0.95) Cathodes for High-Energy-Density Lithium-Ion Batteries: Bulk or Surface Degradation? Chemistry of Materials 30, 1155–1163 (2018). 59. Jung, S. K. et al. Understanding the Degradation Mechanisms of LiNi0.5Co0.2Mn0.3O2 Cathode Material in Lithium Ion Batteries. AdEnM 4, 1300787 (2014). 60. Liu, X. et al. Revealing the surface-to-bulk degradation mechanism of nickel-rich cathode in sulfide all-solid-state batteries. Energy Storage Mater 54, 713–723 (2023). 61. Jung, R., Metzger, M., Maglia, F., Stinner, C. & Gasteiger, H. A. Oxygen Release and Its Effect on the Cycling Stability of LiNi x Mn y Co z O 2 (NMC) Cathode Materials for Li-Ion Batteries . J Electrochem Soc 164, A1361–A1377 (2017). 62. Kondrakov, A. O. et al. Anisotropic Lattice Strain and Mechanical Degradation of High- and Low-Nickel NCM Cathode Materials for Li-Ion Batteries. Journal of Physical Chemistry C 121, 3286–3294 (2017). 63. Li, W., Reimers, J. N. & Dahn, J. R. In situ x-ray diffraction and electrochemical studies of Li1−xNiO2. Solid State Ion 67, 123–130 (1993). 188 References 64. Noh, H. J., Youn, S., Yoon, C. S. & Sun, Y. K. Comparison of the structural and electrochemical properties of layered Li[NixCoyMnz]O2 (x = 1/3, 0.5, 0.6, 0.7, 0.8 and 0.85) cathode material for lithium-ion batteries. J Power Sources 233, 121–130 (2013). 65. Heenan, T. M. M. et al. Identifying the Origins of Microstructural Defects Such as Cracking within Ni-Rich NMC811 Cathode Particles for Lithium-Ion Batteries. Adv Energy Mater 10, 2002655 (2020). 66. Liu, H. et al. Intergranular Cracking as a Major Cause of Long-Term Capacity Fading of Layered Cathodes. Nano Lett 17, 3452–3457 (2017). 67. Schweidler, S. et al. Investigation into Mechanical Degradation and Fatigue of High-Ni NCM Cathode Material: A Long-Term Cycling Study of Full Cells. ACS Appl Energy Mater 2, 7375–7384 (2019). 68. Yan, P. et al. Intragranular cracking as a critical barrier for high-voltage usage of layer- structured cathode for lithium-ion batteries. Nat Commun 8, 14101 (2017). 69. Min, K. & Cho, E. Intrinsic origin of intra-granular cracking in Ni-rich layered oxide cathode materials. Physical Chemistry Chemical Physics 20, 9045–9052 (2018). 70. Tian, H., Gao, L. T. & Guo, Z.-S. Microstructural Adjusting Crack Evolution of Polycrystalline NCM Particle during Charge/Discharge Cycle. J Electrochem Soc 169, 090513 (2022). 71. Cha, H. et al. Boosting Reaction Homogeneity in High-Energy Lithium-Ion Battery Cathode Materials. Advanced Materials 32, 2003040 (2020). 72. Gu, M. et al. Conflicting roles of nickel in controlling cathode performance in lithium ion batteries. Nano Lett 12, 5186–5191 (2012). 73. Zhao, X. et al. Structure Design and Performance of LiNixCoyMn1-x-yO2 Cathode Materials for Lithium-ion Batteries: A Review. Journal of the Chinese Chemical Society 61, 1071–1083 (2014). 74. Shannon, R. D. Revised effective ionic radii and systematic studies of interatomic distances in halides and chalcogenides. Acta Cryst 32, 751–767 (1976). 75. Zhang, M. J. et al. Cationic Ordering Coupled to Reconstruction of Basic Building Units during Synthesis of High-Ni Layered Oxides. J Am Chem Soc 140, 12484–12492 (2018). 76. Liu, W. et al. Nickel-Rich Layered Lithium Transition-Metal Oxide for High-Energy Lithium-Ion Batteries. Angewandte Chemie International Edition 54, 4440–4457 (2015). References 189 77. Lin, F. et al. Surface reconstruction and chemical evolution of stoichiometric layered cathode materials for lithium-ion batteries. Nature Communications 2014 5:1 5, 1–9 (2014). 78. Xu, C. et al. Bulk fatigue induced by surface reconstruction in layered Ni-rich cathodes for Li-ion batteries. Nat Mater 20, 84–92 (2021). 79. Kleiner, K. et al. Fatigue of LiNi0.8Co0.15Al0.05O2 in commercial Li ion batteries. J Power Sources 273, 70–82 (2015). 80. Zheng, J. et al. Corrosion/fragmentation of layered composite cathode and related capacity/voltage fading during cycling process. Nano Lett 13, 3824–3830 (2013). 81. Zhang, F. et al. Surface regulation enables high stability of single-crystal lithium-ion cathodes at high voltage. Nat Commun 11, 3050 (2020). 82. Reed, J. & Ceder, G. Role of electronic structure in the susceptibility of metastable transition-metal oxide structures to transformation. Chem Rev 104, 4513–4533 (2004). 83. Radin, M. D. et al. Narrowing the Gap between Theoretical and Practical Capacities in Li-Ion Layered Oxide Cathode Materials. Adv Energy Mater 7, 1602888 (2017). 84. Liu, T. et al. Rational design of mechanically robust Ni-rich cathode materials via concentration gradient strategy. Nat Commun 12, 6024 (2021). 85. Li, S. et al. Mutual modulation between surface chemistry and bulk microstructure within secondary particles of nickel-rich layered oxides. Nat Commun 11, 4433 (2020). 86. Park, H. et al. In situ multiscale probing of the synthesis of a Ni-rich layered oxide cathode reveals reaction heterogeneity driven by competing kinetic pathways. Nature Chemistry 2022 14:6 14, 614–622 (2022). 87. Wu, Y. et al. Performance and Stability Improvement of Layered NCM Lithium-Ion Batteries at High Voltage by a Microporous Al2O3 Sol-Gel Coating. ACS Omega 4, 13972–13980 (2019). 88. Kim, A. Y. et al. Stabilizing Effect of a Hybrid Surface Coating on a Ni-Rich NCM Cathode Material in All-Solid-State Batteries. Chemistry of Materials 31, 9664–9672 (2019). 89. Guan, P. et al. Recent progress of surface coating on cathode materials for high- performance lithium-ion batteries. Journal of Energy Chemistry 43, 220–235 (2020). 90. Min, K., Seo, S. W., Song, Y. Y., Lee, H. S. & Cho, E. A first-principles study of the preventive effects of Al and Mg doping on the degradation in LiNi0.8Co0.1Mn0.1O2 cathode materials. Physical Chemistry Chemical Physics 19, 1762–1769 (2017). 190 References 91. Lv, Y., Cheng, X., Qiang, W. & Huang, B. Improved electrochemical performances of Ni-rich LiNi0.83Co0.12Mn0.05O2 by Mg-doping. J Power Sources 450, 227718 (2020). 92. Lv, Y., Cheng, X., Qiang, W. & Huang, B. Improved electrochemical performances of Ni-rich LiNi0.83Co0.12Mn0.05O2 by Mg-doping. J Power Sources 450, (2020). 93. Dang, R. et al. The Effect of Elemental Doping on Nickel-Rich NCM Cathode Materials of Lithium Ion Batteries. The Journal of Physical Chemistry C 126, 151–159 (2021). 94. Zhu, C. et al. Synergistic Effect of Microstructure Engineering and Local Crystal Structure Tuning to Improve the Cycling Stability of Ni-Rich Cathodes. ACS Appl Mater Interfaces 13, 48720–48729 (2021). 95. Min, K., Seo, S. W., Song, Y. Y., Lee, H. S. & Cho, E. A first-principles study of the preventive effects of Al and Mg doping on the degradation in LiNi0.8Co0.1Mn0.1O2 cathode materials. Physical Chemistry Chemical Physics 19, 1762–1769 (2017). 96. Kim, U. H. et al. Heuristic solution for achieving long-term cycle stability for Ni-rich layered cathodes at full depth of discharge. Nature Energy 2020 5:11 5, 860–869 (2020). 97. Sun, H. H. et al. Beyond Doping and Coating: Prospective Strategies for Stable High- Capacity Layered Ni-Rich Cathodes. ACS Energy Lett 5, 1136–1146 (2020). 98. Bi, Y. et al. Highly stable Ni-rich layered oxide cathode enabled by a thick protective layer with bio-tissue structure. Energy Storage Mater 24, 291–296 (2020). 99. Liang, L. et al. Surface/Interface Structure Degradation of Ni-Rich Layered Oxide Cathodes toward Lithium-Ion Batteries: Fundamental Mechanisms and Remedying Strategies. Adv Mater Interfaces 7, 1901749 (2020). 100. Huang, H. et al. Pulse High Temperature Sintering to Prepare Single-Crystal High Nickel Oxide Cathodes with Enhanced Electrochemical Performance. Adv Energy Mater 13, (2023). 101. Qian, G. et al. Single-crystal nickel-rich layered-oxide battery cathode materials: synthesis, electrochemistry, and intra-granular fracture. Energy Storage Mater 27, 140– 149 (2020). 102. Su, Y. et al. Stress accumulation in Ni-rich layered oxide cathodes: Origin, impact, and resolution. Journal of Energy Chemistry 65, 236–253 (2022). 103. Mesnier, A. & Manthiram, A. Heuristics for Molten-Salt Synthesis of Single-Crystalline Ultrahigh-Nickel Layered Oxide Cathodes. ACS Appl Mater Interfaces 15, 12895–12907 (2023). References 191 104. Langdon, J. & Manthiram, A. A perspective on single-crystal layered oxide cathodes for lithium-ion batteries. Energy Storage Mater 37, 143–160 (2021). 105. Pang, P. et al. Crack-free single-crystal LiNi0.83Co0.10Mn0.07O2 as cycling/thermal stable cathode materials for high-voltage lithium-ion batteries. Electrochim Acta 365, 137380 (2021). 106. Trevisanello, E., Ruess, R., Conforto, G., Richter, F. H. & Janek, J. Polycrystalline and Single Crystalline NCM Cathode Materials—Quantifying Particle Cracking, Active Surface Area, and Lithium Diffusion. Adv Energy Mater 11, (2021). 107. Deng, X. et al. A Comparative Investigation of Single Crystal and Polycrystalline Ni- Rich NCMs as Cathodes for Lithium-Ion Batteries. Energy and Environmental Materials 6, (2023). 108. Stallard, J. C. et al. Mechanical properties of cathode materials for lithium-ion batteries. Joule 6, 984–1007 (2022). 109. Huang, H. et al. Unusual double ligand holes as catalytic active sites in LiNiO2. Nature Communications 2023 14:1 14, 2112- (2023). 110. Lee, S., Su, L., Mesnier, A., Cui, Z. & Manthiram, A. Cracking vs. surface reactivity in high-nickel cathodes for lithium-ion batteries. Joule 7, 2430–2444 (2023). 111. Moiseev, I. A. et al. Single crystal Ni-rich NMC cathode materials for lithium-ion batteries with ultra-high volumetric energy density. Energy Advances 1, 677–681 (2022). 112. Langdon, J. & Manthiram, A. A perspective on single-crystal layered oxide cathodes for lithium-ion batteries. Energy Storage Mater 37, 143–160 (2021). 113. Wang, Y., Wang, E., Zhang, X. & Yu, H. High-Voltage ‘single-Crystal’ Cathode Materials for Lithium-Ion Batteries. Energy and Fuels 35, 1918–1932 (2021). 114. Raju, K. et al. Influence of Cathode Calendering Density on the Cycling Stability of Li- Ion Batteries Using NMC811 Single or Poly Crystalline Particles. J Electrochem Soc 171, 080519 (2024). 115. Hawley, W. B., Li, M. & Li, J. Room-Temperature Eutectic Synthesis for Upcycling of Cathode Materials. Batteries 9, (2023). 116. Hillier, J. & Baker, R. F. Microanalysis by Means of Electrons. J Appl Phys 15, 663–675 (1944). 192 References 117. Tan, H., Turner, S., Yücelen, E., Verbeeck, J. & Van Tendeloo, G. 2D atomic mapping of oxidation states in transition metal oxides by scanning transmission electron microscopy and electron energy-loss spectroscopy. Phys Rev Lett 107, 107602 (2011). 118. Ramasse, Q. M. et al. Probing the bonding and electronic structure of single atom dopants in graphene with electron energy loss spectroscopy. Nano Lett 13, 4989–4995 (2012). 119. Digital Micrograph, from the Gatan corporate webpage. https://www.gatan.com/products/tem-analysis/digitalmicrograph-software. 120. Peña, F. de la et al. hyperspy/hyperspy: v2.3.0. https://hyperspy.org/hyperspy- doc/v1.6/user_guide/eels.html doi:10.5281/ZENODO.14956374. 121. Yedra, L. et al. Oxide Wizard: an EELS application to characterize the white lines of transition metal edges. Microsc Microanal 20, 698–705 (2014). 122. NumPy. https://numpy.org/. 123. SciPy. https://scipy.org/. 124. Egerton, R. F. Electron Energy-Loss Spectroscopy in the Electron Microscope. Electron Energy-Loss Spectroscopy in the Electron Microscope (Springer US, 2011). doi:10.1007/978- 1-4419-9583-4. 125. Gleason, S. P., Lu, D. & Ciston, J. Prediction of the Cu oxidation state from EELS and XAS spectra using supervised machine learning. npj Computational Materials 2024 10:1 10, 1–10 (2024). 126. Graetz, J., Ahn, C. C., Ouyang, H., Rez, P. & Fultz, B. White lines and d-band occupancy for the 3d transition-metal oxides and lithium transition-metal oxides. Phys Rev B Condens Matter Mater Phys 69, 235103-1-235103–6 (2004). 127. Van Aken, P. A., Liebscher, B. & Styrsa, V. J. Quantitative determination of iron oxidation states in minerals using Fe L2,3-edge electron energy-loss near-edge structure spectroscopy. Phys Chem Miner 25, 323–327 (1998). 128. Varela, M. et al. Atomic-resolution imaging of oxidation states in manganites. Phys Rev B Condens Matter Mater Phys 79, 085117 (2009). 129. Cavé, L., Al, T., Loomer, D., Cogswell, S. & Weaver, L. A STEM/EELS method for mapping iron valence ratios in oxide minerals. Micron 37, 301–309 (2006). References 193 130. Tan, H., Verbeeck, J., Abakumov, A. & Van Tendeloo, G. Oxidation state and chemical shift investigation in transition metal oxides by EELS. Ultramicroscopy 116, 24–33 (2012). 131. del-Pozo-Bueno, D., Peiró, F. & Estradé, S. Support vector machine for EELS oxidation state determination. Ultramicroscopy 221, 113190 (2021). 132. Del-Pozo-Bueno, D. et al. Direct Evidence of a Graded Magnetic Interface in Bimagnetic Core/Shell Nanoparticles Using Electron Magnetic Circular Dichroism (EMCD). Nano Lett 21, 6923–6930 (2021). 133. Torruella, P. et al. Clustering analysis strategies for electron energy loss spectroscopy (EELS). Ultramicroscopy 185, 42–48 (2018). 134. Potapov, P. Why Principal Component Analysis of STEM spectrum-images results in “abstract”, uninterpretable loadings? Ultramicroscopy 160, 197–212 (2016). 135. Lichtert, S. & Verbeeck, J. Statistical consequences of applying a PCA noise filter on EELS spectrum images. Ultramicroscopy 125, 35–42 (2013). 136. Cueva, P., Hovden, R., Mundy, J. A., Xin, H. L. & Muller, D. A. New Approaches to Data Processing for Atomic Resolution EELS. Microscopy and Microanalysis 18, 970–971 (2012). 137. Mousavi M, S. S., Pofelski, A., Teimoori, H. & Botton, G. A. Alignment-invariant signal reality reconstruction in hyperspectral imaging using a deep convolutional neural network architecture. Sci Rep 12, (2022). 138. Annys, A., Jannis, D. & Verbeeck, J. Deep learning for automated materials characterisation in core-loss electron energy loss spectroscopy. Sci Rep 13, 1–10 (2023). 139. Chen, H., Nabiei, F., Badro, J., Alexander, D. T. & Hébert, C. Machine Learning on STEM-EDS Data for Quantifying Overlapping Deep-Mantle Rock Assemblages. Microscopy and Microanalysis 26, 1878–1880 (2020). 140. del-Pozo-Bueno, D., Peiró, F. & Estradé, S. Support vector machine for EELS oxidation state determination. Ultramicroscopy 221, 113190 (2021). 141. Ji, Z., Hu, M. & Xin, H. L. MnEdgeNet for accurate decomposition of mixed oxidation states for Mn XAS and EELS L2,3 edges without reference and calibration. Scientific Reports 2023 13:1 13, 1–14 (2023). 142. Vetter, J. et al. Ageing mechanisms in lithium-ion batteries. J Power Sources 147, 269–281 (2005). 194 References 143. Fergus, J. W. Recent developments in cathode materials for lithium ion batteries. J Power Sources 195, 939–954 (2010). 144. Wiedemann, A. H., Goldin, G. M., Barnett, S. A., Zhu, H. & Kee, R. J. Effects of three- dimensional cathode microstructure on the performance of lithium-ion battery cathodes. Electrochim Acta 88, 580–588 (2013). 145. Malavé, V., Berger, J. R., Zhu, H. & Kee, R. J. A Computational Model of the Mechanical Behavior within Reconstructed LixCoO2 Li-ion Battery Cathode Particles. Electrochim Acta 130, 707–717 (2014). 146. Goldin, G. M., Colclasure, A. M., Wiedemann, A. H. & Kee, R. J. Three-dimensional particle-resolved models of Li-ion batteries to assist the evaluation of empirical parameters in one-dimensional models. Electrochim Acta 64, 118–129 (2012). 147. Hutzenlaub, T., Thiele, S., Zengerle, R. & Ziegler, C. Three-dimensional reconstruction of a LiCoO 2 Li-ion battery cathode. Electrochemical and Solid-State Letters 15, A33 (2012). 148. Ender, M., Joos, J., Carraro, T. & Ivers-Tiffée, E. Quantitative Characterization of LiFePO 4 Cathodes Reconstructed by FIB/SEM Tomography . J Electrochem Soc 159, A972–A980 (2012). 149. Ender, M., Joos, J., Weber, A. & Ivers-Tiffée, E. Anode microstructures from high-energy and high-power lithium-ion cylindrical cells obtained by X-ray nano-tomography. J Power Sources 269, 912–919 (2014). 150. Dragonfly 2024.1. https://dragonfly.comet.tech/. 151. Wiedenmann, D. et al. Three-dimensional pore structure and ion conductivity of porous ceramic diaphragms. AIChE Journal 59, 1446–1457 (2013). 152. Avizo Software | Materials Characterization Software - UK. https://www.thermofisher.com/uk/en/home/electron-microscopy/products/software- em-3d-vis/avizo-software.html. 153. Cooper, S. J., Bertei, A., Shearing, P. R., Kilner, J. A. & Brandon, N. P. TauFactor: An open-source application for calculating tortuosity factors from tomographic data. SoftwareX 5, 203–210 (2016). 154. Tan, C., Kok, M. D. R., Daemi, S. R., Brett, D. J. L. & Shearing, P. R. Three-dimensional image based modelling of transport parameters in lithium–sulfur batteries. Physical Chemistry Chemical Physics 21, 4145–4154 (2019). References 195 155. Newman, J. & Chapman, T. W. Restricted diffusion in binary solutions. AIChE Journal 19, 343–348 (1973). 156. Harned, H. S. & French, D. M. A Conductance Method for the Determination of the Diffusion Coefficients of Electrolytes. Ann N Y Acad Sci 46, 267–284 (1945). 157. Landesfeind, J., Hattendorff, J., Ehrl, A., Wall, W. A. & Gasteiger, H. A. Tortuosity Determination of Battery Electrodes and Separators by Impedance Spectroscopy. J Electrochem Soc 163, A1373–A1387 (2016). 158. Malifarge, S., Delobel, B. & Delacourt, C. Determination of Tortuosity Using Impedance Spectra Analysis of Symmetric Cell. J Electrochem Soc 164, E3329–E3334 (2017). 159. Williams, D. B. & Carter, C. B. Transmission electron microscopy: A textbook for materials science. Transmission Electron Microscopy: A Textbook for Materials Science 1–760 (2009) doi:10.1007/978-0-387-76501-3. 160. Baliyan, A., Imai, H. & Kumar, V. Microscopy. Data Processing Handbook for Complex Biological Data Sources 97–117 (2019) doi:10.1016/B978-0-12-816548-5.00007-1. 161. Wang, Z. Lin. Elastic and Inelastic Scattering in Electron Diffraction and Imaging. (Plenum Press, 1995). 162. Rose, H. H. Optics of high-performance electron microscopes. Sci Technol Adv Mater 9, 014107 (2016). 163. Watanabe, M. & Egerton, R. F. Fundamental limit to single-atom analysis by STEM- EDX spectroscopy. Microscopy and Microanalysis 23, 380–381 (2017). 164. Pennycook, S. J. Seeing the atoms more clearly: STEM imaging from the Crewe era to today. Ultramicroscopy 123, 28–37 (2012). 165. von Ardenne, M. Das Elektronen-Rastermikroskop - Theoretische Grundlagen. Zeitschrift für Physik 109, 553–572 (1938). 166. Crewe, A. V. High Resolution Scanning Microscopy of Biological Specimens. Phil. Trans. Roy. Soc. Lond. B vol. 261 https://royalsocietypublishing.org/ (1971). 167. Nellist, P. D. & Pennycook, S. J. The principles and interpretation of annular dark-field Z-contrast imaging. Advances in Imaging and Electron Physics 113, 147–203 (2000). 168. Browning, N. D. et al. The Application of Scanning Transmission Electron Microscopy (STEM) to the Study of Nanoscale Systems. 11–40 (2012) doi:10.1007/978-1-4614-2191- 7_2. 196 References 169. Nellist, P. D. The Principles of STEM Imaging. Scanning Transmission Electron Microscopy 91–115 (2011) doi:10.1007/978-1-4419-7200-2_2. 170. Radamson, H. H. Electron Microscopy. 115–146 (2023) doi:10.1007/978-3-031-26434-4_4. 171. Locate Elements within Sample | EELS.info. https://eels.info/uses/locate-elements- within-sample. 172. The influence of electrochemical cycling protocols on capacity loss in nickel-rich lithium-ion batteries - Journal of Materials Chemistry A (RSC Publishing) DOI:10.1039/D1TA06324C. https://pubs.rsc.org/en/content/articlehtml/2021/ta/d1ta06324c. 173. Langford, R. M. & Clinton, C. In situ lift-out using a FIB-SEM system. Micron 35, 607– 611 (2004). 174. Pease, R. F. W. Significant Advances in Scanning Electron Microscopes (1965–2007). Advances in Imaging and Electron Physics 150, 53–86 (2008). 175. Munroe, P. R. The application of focused ion beam microscopy in the material sciences. Mater Charact 60, 2–13 (2009). 176. FIB Lift Out Grids - Copper. https://www.agarscientific.com/fib-lift-out-grids-copper. 177. Liu, Y. et al. Nano-Tomography of Porous Geological Materials Using Focused Ion Beam-Scanning Electron Microscopy. Minerals 2016, Vol. 6, Page 104 6, 104 (2016). 178. Xlib. https://imagej.net/plugins/xlib/. 179. Schmuch, R., Wagner, R., Hörpel, G., Placke, T. & Winter, M. Performance and cost of materials for lithium-based rechargeable automotive batteries. Nature Energy 2018 3:4 3, 267–278 (2018). 180. Cano, Z. P. et al. Batteries and fuel cells for emerging electric vehicle markets. Nature Energy 2018 3:4 3, 279–289 (2018). 181. Daemi, S. R. et al. Visualizing the Carbon Binder Phase of Battery Electrodes in Three Dimensions. ACS Appl Energy Mater 1, 3702–3710 (2018). 182. Ge, R. et al. Numerical Design of Microporous Carbon Binder Domains Phase in Composite Cathodes for Lithium-Ion Batteries. ACS Appl Mater Interfaces 15, 27809– 27820 (2023). References 197 183. Wang, Y., Fu, X., Zheng, M., Zhong, W. H. & Cao, G. Strategies for Building Robust Traffic Networks in Advanced Energy Storage Devices: A Focus on Composite Electrodes. Advanced Materials 31, (2019). 184. Park, J. et al. Challenges, laser processing and electrochemical characteristics on application of ultra-thick electrode for high-energy lithium-ion battery. J Power Sources 482, 228948 (2021). 185. Kim, W. et al. Lithium-selenium sulfide batteries with long cycle life and high energy density via solvent washing treatment. Appl Surf Sci 512, 145632 (2020). 186. Kim, W. et al. Advanced ordered mesoporous carbon with fast Li-ion diffusion for lithium–selenium sulfide batteries in a carbonate-based electrolyte. Carbon N Y 170, 236–244 (2020). 187. Vogel, J. E. et al. Electrode microstructure controls localized electronic impedance in Li- ion batteries. Electrochim Acta 297, 820–825 (2019). 188. Kondo, H. & Srinivasan, V. Simulation study of rate limiting factors of Li-ion batteries using experimental functions of electronic and ionic resistances. Electrochim Acta 371, 137834 (2021). 189. Zheng, H., Tan, L., Liu, G., Song, X. & Battaglia, V. S. Calendering effects on the physical and electrochemical properties of Li[Ni1/3Mn1/3Co1/3]O2 cathode. J Power Sources 208, 52–57 (2012). 190. Primo, E. N., Touzin, M. & Franco, A. A. Calendering of Li(Ni0.33Mn0.33Co0.33)O2- Based Cathodes: Analyzing the Link Between Process Parameters and Electrode Properties by Advanced Statistics. Batter Supercaps 4, 834–844 (2021). 191. Rucci, A., Ngandjong, A. C., Primo, E. N., Maiza, M. & Franco, A. A. Tracking variabilities in the simulation of Lithium Ion Battery electrode fabrication and its impact on electrochemical performance. Electrochim Acta 312, 168–178 (2019). 192. Duffner, F. et al. Post-lithium-ion battery cell production and its compatibility with lithium-ion cell production infrastructure. Nat Energy 6, 123–134 (2021). 193. Trembacki, B. L., Noble, D. R., Ferraro, M. E. & Roberts, S. A. Mesoscale Effects of Composition and Calendering in Lithium-Ion Battery Composite Electrodes. Journal of Electrochemical Energy Conversion and Storage 17, (2020). 194. M Heenan, T. M. et al. Identifying the Origins of Microstructural Defects Such as Cracking within Ni-Rich NMC811 Cathode Particles for Lithium-Ion Batteries. Adv Energy Mater 10, 2002655 (2020). 198 References 195. Schreiner, D., Oguntke, M., Günther, T. & Reinhart, G. Modelling of the Calendering Process of NMC-622 Cathodes in Battery Production Analyzing Machine/Material– Process–Structure Correlations. Energy Technology 7, 1900840 (2019). 196. Lu, X. et al. Microstructural Evolution of Battery Electrodes During Calendering. Joule 4, 2746–2768 (2020). 197. Lee, S. G. & Jeon, D. H. Effect of electrode compression on the wettability of lithium- ion batteries. J Power Sources 265, 363–369 (2014). 198. Cha, H. et al. Boosting Reaction Homogeneity in High-Energy Lithium-Ion Battery Cathode Materials. Advanced Materials 32, (2020). 199. Kang, H. et al. Geometric and Electrochemical Characteristics of LiNi1/3Mn1/3Co1/3O2 Electrode with Different Calendering Conditions. Electrochim Acta 232, 431–438 (2017). 200. Lee, S. et al. In-Depth Analysis of the Degradation Mechanisms of High-Nickel, Low/No-Cobalt Layered Oxide Cathodes for Lithium-Ion Batteries. Adv Energy Mater 11, 2100858 (2021). 201. Schmidt, D., Kamlah, M. & Knoblauch, V. Highly densified NCM-cathodes for high energy Li-ion batteries: Microstructural evolution during densification and its influence on the performance of the electrodes. J Energy Storage 17, 213–223 (2018). 202. Morzy, J. K. et al. Origins and Importance of Intragranular Cracking in Layered Lithium Transition Metal Oxide Cathodes. ACS Appl Energy Mater 7, 3945–3956 (2024). 203. Lee, S., Su, L., Mesnier, A., Cui, Z. & Manthiram, A. Cracking vs. surface reactivity in high-nickel cathodes for lithium-ion batteries. Joule 7, 2430–2444 (2023). 204. Kim, W., Jang, D. & Kim, H. J. Understanding electronic and Li-ion transport of LiNi0.5Co0.2Mn0.3O2 electrodes affected by porosity and electrolytes using electrochemical impedance spectroscopy. J Power Sources 510, 230338 (2021). 205. Usseglio-Viretta, F. L. E. et al. Resolving the Discrepancy in Tortuosity Factor Estimation for Li-Ion Battery Electrodes through Micro-Macro Modeling and Experiment. J Electrochem Soc 165, A3403–A3426 (2018). 206. Daniel, C. & Besenhard, J. O. Handbook of Battery Materials: Second Edition. Handbook of Battery Materials: Second Edition https://doi.org/10.1002/9783527637188 (2011) doi:10.1002/9783527637188. References 199 207. Ge, R., Cumming, D. J. & Smith, R. M. Discrete element method (DEM) analysis of lithium ion battery electrode structures from X-ray tomography-the effect of calendering conditions. Powder Technol 403, 117366 (2022). 208. Patel, K. K., Paulsen, J. M. & Desilvestro, J. Numerical simulation of porous networks in relation to battery electrodes and separators. J Power Sources 122, 144–152 (2003). 209. Bruggeman, D. A. G. Berechnung verschiedener physikalischer Konstanten von heterogenen Substanzen. I. Dielektrizitätskonstanten und Leitfähigkeiten der Mischkörper aus isotropen Substanzen. Ann Phys 416, 636–664 (1935). 210. Chung, D.-W., Ebner, M., Ely, D. R., Wood, V. & García, R. E. Validity of the Bruggeman relation for porous electrodes. Model Simul Mat Sci Eng 21, 074009 (2013). 211. Tjaden, B., Cooper, S. J., Brett, D. J., Kramer, D. & Shearing, P. R. On the origin and application of the Bruggeman correlation for analysing transport phenomena in electrochemical systems. Curr Opin Chem Eng 12, 44–51 (2016). 212. Kehrwald, D., Shearing, P. R., Brandon, N. P., Sinha, P. K. & Harris, S. J. Local Tortuosity Inhomogeneities in a Lithium Battery Composite Electrode. J Electrochem Soc 158, A1393 (2011). 213. Cantoni, M. & Holzer, L. Advances in 3D focused ion beam tomography. MRS Bull 39, 354–360 (2014). 214. Jangde, A., Kumar, M., Tuğrul Gülenç, İ., Wheatcroft, L. & Inkson, B. J. Mechanical Properties of Cycled Single Crystal LiNi0.8Mn0.1Co0.1O2 (NMC811) Particles. Batter Supercaps 8, (2025). 215. Fakhrudin, M., Kartini, E. & Zulfia, A. Improving performance of cathode NMC-811 by CeO2-coating for Li-ion battery. Microelectron Eng 288, 112169 (2024). 216. Dose, W. M., Xu, C., Grey, C. P. & De Volder, M. F. L. Effect of Anode Slippage on Cathode Cutoff Potential and Degradation Mechanisms in Ni-Rich Li-Ion Batteries. Cell Rep Phys Sci 1, 100253 (2020). 217. Dijkstra, E. W. A note on two problems in connexion with graphs. Numer Math (Heidelb) 1, 269–271 (1959). 218. Kench, S., Squires, I. & Cooper, S. TauFactor 2: A GPU accelerated python tool for microstructural analysis. J Open Source Softw 8, 5358 (2023). 219. Newman, J. & Balsara, N. P. Electrochemical Systems. (The Electrochemical Society, 2021). 200 References 220. Mulder, W. H., Sluyters, J. H., Pajkossy, T. & Nyikos, L. Tafel current at fractal electrodes: Connection with admittance spectra. J Electroanal Chem Interfacial Electrochem 285, 103–115 (1990). 221. Landesfeind, J., Pritzl, D. & Gasteiger, H. A. An Analysis Protocol for Three-Electrode Li-Ion Battery Impedance Spectra: Part I. Analysis of a High-Voltage Positive Electrode. J Electrochem Soc 164, A1773–A1783 (2017). 222. Wang, C.-W., Sastry, A. M., Striebel, K. A. & Zaghib, K. Extraction of Layerwise Conductivities in Carbon-Enhanced, Multilayered LiFePO[sub 4] Cathodes. J Electrochem Soc 152, A1001 (2005). 223. Ender, M., Weber, A. & Ivers-Tiffée, E. A novel method for measuring the effective conductivity and the contact resistance of porous electrodes for lithium-ion batteries. Electrochem commun 34, 130–133 (2013). 224. Ogihara, N., Itou, Y., Sasaki, T. & Takeuchi, Y. Impedance spectroscopy characterization of porous electrodes under different electrode thickness using a symmetric cell for high-performance lithium-ion batteries. Journal of Physical Chemistry C 119, 4612–4619 (2015). 225. Liu, Y. et al. Proton Conduction and Oxygen Reduction Kinetics in PEM Fuel Cell Cathodes: Effects of Ionomer-to-Carbon Ratio and Relative Humidity. J Electrochem Soc 156, B970 (2009). 226. Heenan, T. M. M. et al. Identifying the Origins of Microstructural Defects Such as Cracking within Ni-Rich NMC811 Cathode Particles for Lithium-Ion Batteries. Adv Energy Mater 10, (2020). 227. Yan, P. et al. Intragranular cracking as a critical barrier for high-voltage usage of layer- structured cathode for lithium-ion batteries. Nature Communications 2017 8:1 8, 1–9 (2017). 228. Lin, Q. et al. Ni–Li anti-site defect induced intragranular cracking in Ni-rich layer- structured cathode. Nano Energy 76, 105021 (2020). 229. Zhang, H. et al. Rock-Salt Growth-Induced (003) Cracking in a Layered Positive Electrode for Li-Ion Batteries. ACS Energy Lett 2, 2607–2615 (2017). 230. Merryweather, A. J., Schnedermann, C., Jacquet, Q., Grey, C. P. & Rao, A. Operando optical tracking of single-particle ion dynamics in batteries. Nature 2021 594:7864 594, 522–528 (2021). References 201 231. Xu, C. et al. Bulk fatigue induced by surface reconstruction in layered Ni-rich cathodes for Li-ion batteries. Nature Materials 2020 20:1 20, 84–92 (2020). 232. Mao, Y. et al. High-Voltage Charging-Induced Strain, Heterogeneity, and Micro-Cracks in Secondary Particles of a Nickel-Rich Layered Cathode Material. Adv Funct Mater 29, 1900247 (2019). 233. Parks, H. C. W. et al. Direct observations of electrochemically induced intergranular cracking in polycrystalline NMC811 particles. J Mater Chem A Mater 11, 21322–21332 (2023). 234. Duan, S. et al. Three-dimensional reconstruction and computational analysis of a structural battery composite electrolyte. Commun Mater 4, (2023). 235. Vijayaraghavan, B., Ely, D. R., Chiang, Y.-M., García-García, R. & García, R. E. An Analytical Method to Determine Tortuosity in Rechargeable Battery Electrodes. J Electrochem Soc 159, A548–A552 (2012). 236. Dong, H. & Blunt, M. J. Pore-network extraction from micro-computerized- tomography images. https://doi.org/10.1103/PhysRevE.80.036307 doi:10.1103/PhysRevE.80.036307. 237. Gostick, J. T. Versatile and efficient pore network extraction method using marker- based watershed segmentation. Phys Rev E 96, 023307 (2017). 238. Blunt, M. J. et al. Pore-scale imaging and modelling. Adv Water Resour 51, 197–216 (2013). 239. Pritzl, D. et al. Identifying Contact Resistances in High-Voltage Cathodes by Impedance Spectroscopy. J Electrochem Soc 166, A582–A590 (2019). 240. Sim, R., Lee, S., Li, W. & Manthiram, A. Influence of Calendering on the Electrochemical Performance of LiNi0.9Mn0.05Al0.05O2Cathodes in Lithium-Ion Cells. ACS Appl Mater Interfaces 13, 42898–42908 (2021). 241. Ebner, M., Chung, D. W., García, R. E. & Wood, V. Tortuosity anisotropy in lithium-ion battery electrodes. Adv Energy Mater 4, (2014). 242. Ryu, H. H., Lee, S. B., Yoon, C. S. & Sun, Y. K. Morphology-Dependent Battery Performance of Ni-Rich Layered Cathodes: Single-Crystal versus Refined Polycrystal. ACS Energy Lett 7, 3072–3079 (2022). 243. Doyle, M., Newman, J., Gozdz, A. S., Schmutz, C. N. & Tarascon, J. Comparison of Modeling Predictions with Experimental Data from Plastic Lithium Ion Cells. J Electrochem Soc 143, 1890–1903 (1996). 202 References 244. Cooper, S. J. et al. Image based modelling of microstructural heterogeneity in LiFePO4 electrodes for Li-ion batteries. J Power Sources 247, 1033–1039 (2014). 245. Stephenson, D. E., Hartman, E. M., Harb, J. N. & Wheeler, D. R. Modeling of Particle- Particle Interactions in Porous Cathodes for Lithium-Ion Batteries. J Electrochem Soc 154, A1146 (2007). 246. Alolaywi, H. Y., Uzun, K. & Cheng, Y. T. “Zero” Porosity High Loading NMC622 Positive Electrodes for Li-Ion Batteries. J Electrochem Soc 171, 010514 (2024). 247. Alolaywi, H., Uzun, K. & Cheng, Y. T. Low porosity NMC622 and NMC811 electrodes made by severe calendering. J Energy Storage 105, 114559 (2025). 248. Tawfik, M. I., Ali, A. & Asfoor, M. A combined trade-off strategy of battery degradation, charge retention, and driveability for electric vehicles. Sci Rep 14, 21995 (2024). 249. Sayed, K., Aref, M., Almalki, M. M. & Mossa, M. A. Optimizing fast charging protocols for lithium-ion batteries using reinforcement learning: Balancing speed, efficiency, and longevity. Results in Engineering 25, 104302 (2025). 250. Tebbe, J. et al. Innovations and prognostics in battery degradation and longevity for energy storage systems. J Energy Storage 114, 115724 (2025). 251. Dimitrijevic, A. The influence of electrochemical history on lifetime of commercial lithium-ion batteries. Doctoral thesis, UCL (University College London). (2024). 252. Quilty, C. D. et al. Elucidating Cathode Degradation Mechanisms in LiNi0.8Mn0.1Co0.1O2 (NMC811)/Graphite Cells Under Fast Charge Rates Using Operando Synchrotron Characterization. J Electrochem Soc 169, 020545 (2022). 253. Björklund, E. et al. Cycle-Induced Interfacial Degradation and Transition-Metal Cross- Over in LiNi0.8Mn0.1Co0.1O2-Graphite Cells. Chemistry of Materials 34, 2034–2048 (2022). 254. Pennycook, S. J. The impact of STEM aberration correction on materials science. Ultramicroscopy 180, 22–33 (2017). 255. Alexander, J. A. et al. High-resolution monochromated electron energy-loss spectroscopy of organic photovoltaic materials. Ultramicroscopy 180, 125–132 (2017). 256. Hachtel, J. A. et al. Exploring the capabilities of monochromated electron energy-loss spectroscopy in the infrared regime. Scientific Reports 8, 5637 (2018). References 203 257. Krivanek, O. L. et al. Progress in ultrahigh energy resolution EELS. Ultramicroscopy 203, 60–67 (2019). 258. Lin, F., Markus, I. M., Doeff, M. M. & Xin, H. L. Chemical and structural stability of lithium-ion battery electrode materials under electron beam. Sci Rep 4, 1–6 (2014). 259. Kim, H. G., Myung, S. T., Lee, J. K. & Sun, Y. K. Effects of manganese and cobalt on the electrochemical and thermal properties of layered Li[Ni0.52Co0.16+xMn0.32−x]O2 cathode materials. J Power Sources 196, 6710–6715 (2011). 260. Blanco-Portals, J. et al. WhatEELS. A python-based interactive software solution for ELNES analysis combining clustering and NLLS. Ultramicroscopy 232, 113403 (2022). 261. Kong, L., Ji, Z. & Xin, H. L. Electron energy loss spectroscopy database synthesis and automation of core-loss edge recognition by deep-learning neural networks. Sci Rep 12, (2022). 262. An, L. et al. Distinguishing bulk redox from near-surface degradation in lithium nickel oxide cathodes. Energy Environ Sci 17, 8379–8391 (2024). 263. Wang, Z. L., Yin, J. S. & Jiang, Y. D. EELS analysis of cation valence states and oxygen vacancies in magnetic oxides. Micron 31, 571–580 (2000). 264. Fan, X. et al. Crack-free single-crystalline Ni-rich layered NCM cathode enable superior cycling performance of lithium-ion batteries. Nano Energy 70, 104450 (2020). 265. Leapman, R. D., Rez, P. & Mayers, D. F. K, L, and M shell generalized oscillator strengths and ionization cross sections for fast electron collisions. J Chem Phys 72, 1232– 1243 (1980). 266. Rez, P. Cross-sections for energy loss spectrometry. Ultramicroscopy 9, 283–287 (1982). 267. Ahn, C. C. & Rez, P. Inner shell edge profiles in electron energy loss spectroscopy. Ultramicroscopy 17, 105–115 (1985). 268. Groot, F. De. Multiplet effects in X-ray spectroscopy. Coord Chem Rev 249, 31–63 (2005). 269. coding4mc/mc-eels. https://github.com/coding4mc/mc-eels. 270. Teichert, P., Eshetu, G. G., Jahnke, H. & Figgemeier, E. Degradation and aging routes of ni-rich cathode based li-ion batteries. Batteries 6, 1–26 (2020). 271. Sparrow, T. G., Williams, B. G., Rao, C. N. R. & Thomas, J. M. L3/L2 white-line intensity ratios in the electron energy-loss spectra of 3d transition-metal oxides. Chem Phys Lett 108, 547–550 (1984). 204 References 272. Graetz, J., Ahn, C. C., Ouyang, H., Rez, P. & Fultz, B. White lines and d-band occupancy for the 3d transition-metal oxides and lithium transition-metal oxides. Phys Rev B Condens Matter Mater Phys 69, 235103-1-235103–6 (2004). 273. Annys, A., Jannis, D. & Verbeeck, J. Deep learning for automated materials characterisation in core-loss electron energy loss spectroscopy. Sci Rep 13, (2023). 274. Pate, C. M., Hart, J. L. & Taheri, M. L. RapidEELS: machine learning for denoising and classification in rapid acquisition electron energy loss spectroscopy. Sci. Rep. 11, 19515 (2021). 275. Varela, M. et al. Atomic-resolution imaging of oxidation states in manganites. Phys Rev B Condens Matter Mater Phys 79, 085117 (2009). 276. Choudhary, K. et al. Recent advances and applications of deep learning methods in materials science. npj Computational Materials 2022 8:1 8, 1–26 (2022). 277. Gu, J. et al. Recent advances in convolutional neural networks. Pattern Recognit 77, 354– 377 (2018). 278. Ioffe, S. & Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. 32nd International Conference on Machine Learning, ICML 2015 1, 448–456 (2015). 279. Kaiming, H., Xiangyu, Z., Shaoqing, R. & Jian, S. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification Kaiming. Biochem Biophys Res Commun 498, 254–261 (2018). 280. Kingma, D. P. & Ba, J. L. Adam: A Method for Stochastic Optimization. 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings https://arxiv.org/pdf/1412.6980 (2014). 281. Mao, A., Mohri, M. & Zhong, Y. Cross-Entropy Loss Functions: Theoretical Analysis and Applications. Proc Mach Learn Res 202, 23803–23828 (2023). 282. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research 15, 1929–1958 (2014). 283. Tian, C., Lin, F. & Doeff, M. M. Electrochemical Characteristics of Layered Transition Metal Oxide Cathode Materials for Lithium Ion Batteries: Surface, Bulk Behavior, and Thermal Properties. Acc. Chem. Res. 51, 89 (2018). References 205 284. Quilty, C. D. et al. Capacity Fade of Graphite/NMC811: Influence of Particle Morphology, Electrolyte, and Charge Voltage. J Electrochem Soc 171, 080515 (2024). 285. Zou, L. et al. Revealing Cycling Rate-Dependent Structure Evolution in Ni-Rich Layered Cathode Materials. ACS Energy Lett 3, 2433–2440 (2018). 286. Chu, Y. et al. Oxygen Release in Ni-Rich Layered Cathode for Lithium-Ion Batteries: Mechanisms and Mitigating Strategies. ChemElectroChem 11, e202300653 (2024). 287. Lee, S., Su, L., Mesnier, A., Cui, Z. & Manthiram, A. Cracking vs. surface reactivity in high-nickel cathodes for lithium-ion batteries. Joule 7, 2430–2444 (2023). 288. Tan, H., Turner, S., Yücelen, E., Verbeeck, J. & Van Tendeloo, G. 2D atomic mapping of oxidation states in transition metal oxides by scanning transmission electron microscopy and electron energy-loss spectroscopy. Phys Rev Lett 107, 107602 (2011). 289. Botton, G. A., Appel, C. C., Horsewell, A. & Stobbs, W. M. Quantification of the EELS near-edge structures to study Mn doping in oxides. J Microsc 180, 211–216 (1995). 290. Egerton, R. F. Electron Energy-Loss Spectroscopy in the Electron Microscope. Electron Energy-Loss Spectroscopy in the Electron Microscope https://doi.org/10.1007/978-1-4419- 9583-4 (2011) doi:10.1007/978-1-4419-9583-4. 291. Chen, J. G. NEXAFS investigations of transition metal oxides, nitrides, carbides, sulfides and other interstitial compounds. Surf Sci Rep 30, 1–152 (1997). 292. Leapman, R. D., Grunes, L. A. & Fejes, P. L. Study of the 𝐿23 edges in the 3⁢𝑑 transition metals and their oxides by electron-energy-loss spectroscopy with comparisons to theory. Phys Rev B 26, 614 (1982). 293. Schmid, H. K. & Mader, W. Oxidation states of Mn and Fe in various compound oxide systems. Micron 37, 426–432 (2006). 206 List of Figures List of Figures Figure 1.1: Global electric car sales by region from 2012 to 2024, showing exponential growth from 0.1 million in 2012 to over 16 million in 2024. For the year 2024, China leads with approximately 10 million sales, followed by Europe (3.4 million), the United States (1.7 million), and the rest of the world (1.4 million). Data source: IEA 3 ............................................. 1 Figure 1.2: Price trends of selected battery materials and lithium-ion battery packs from 2015 to 2024. The index tracks prices of key raw materials including cobalt sulphate, lithium carbonate, nickel sulphate, manganese flake, and phosphoric acid. Data source: IEA12 ........... 3 Figure 2.1: Schematic representation of Li-ion battery operation. Red dots indicate Li-ion movement through the electrolyte, and green arrows show electron flow in the external circuit. The top panel illustrates charging, and the bottom shows discharging. Sizes are not to scale. Figure produced by Morzy.29 ............................................................................................................. 6 Figure 2.2: Historical milestones of the Ni-rich cathode development. The abbreviations NMC consist of LiNi1-x-yMnxCoyO2 where references were taken from NMC11138, NMC42239, NMC532,40 NMC622,41 NMC811,22 NMC91.13 Reference used for LiNixMn1-xO242 and LiNixCo1- xO2 (LiNi0.75Co0.25O2).43 References used for LiCoO2.6,44 Figure produced by Wu et al.45 ........... 10 Figure 2.3: (a) Comparison of Ni4+ concentration of various NCM cathode against different depth of charge. Copyright 2017, American Chemical Society.24 (b) Capacity retention with the increased concentration of Ni. Copyright 2018 Chemistry of Materials.58 ................................. 12 Figure 2.4: (a) Evolution of average NMC811 unit cell volume and lattice parameters (𝑎 and 𝑐) during delithiation measured using X-ray diffraction29,53 (b) dQ/dV profiles of NMC811 between 2.9 and 4.7 V vs. Li/Li+.16 .................................................................................................... 14 Figure 2.5: (a) Ordered 𝑅3𝑚 structure. (b) Cation mixing phase with 𝐹𝑚3𝑚 structure. (c) 𝑅3𝑚 structure with Li vacancies at highly charged state. (d) Cation mixed phase with partial TM ions in Li layer. Liu et al..76 ................................................................................................................ 16 Figure 2.6: The structural evolution during delithiation of the (e) active phase and (f) the fatigued phase. In the illustrations, lithium atoms are represented by green circles, transition List of Figures 207 metal atoms by blue circles, and oxygen atoms by red circles. Figure produced by Kleiner et al..79 ....................................................................................................................................................... 17 Figure 2.7: (a) HAADF-STEM images with the atomic resolution for detection of the surface of pristine NMC samples. The scale bars are 5 nm in left image and 1 nm in region 1 and 2, respectively. (b) Scheme of atomic configuration for bulk and surface in polycrystalline NMC particle. Figure produced by Zhang et al..81 ................................................................................... 18 Figure 2.8: Illustration of the surface-to-bulk reaction coupling effect during battery process. The arrow presents the interaction between the surficial chemistry and the bulk microstructure. Figure produced by Li et al..85 .............................................................................. 19 Figure 2.9: Comparison of single crystal (SC, left) and polycrystalline (PC, right) NMC811 cathodes after calendering. Green ticks and red crosses denote microstructural advantages and disadvantages, respectively. ..................................................................................................... 21 Figure 2.10: (a) High-resolution transmission electron microscopy (HRTEM) image showing a cycled NMC811 particle with a reconstructed rock-salt surface layer (~12 nm thick) overlying the original layered structure. The insets show corresponding fast Fourier transform (FFT) patterns identifying the rock-salt (top) and layered (bottom) phases. (b) HAADF-STEM image of the same region with coloured boxes indicating EELS line-scan positions from surface to bulk. (c) EELS spectra of the O K-edge, Mn L-edge, Co L-edge, and Ni L-edge collected along the scan direction, revealing chemical variations with depth. Figure produced by Li et al.. 27 .............................................................................................................................................................. 24 Figure 2.11: (a) HAADF-STEM image of a cycled NMC811 particle after 20 cycles between 2.0 - 4.8 V at C/10 rate. (b) Corresponding EELS spectra showing the Mn, Co, and Ni L2,3 edges and O K-edge from the surface towards the bulk. (c) Evolution of L3/L2 intensity ratios for Mn, Co, and Ni as a function of depth. Figure produced by Hwang et al..26 .................................... 25 Figure 2.12: Timeline showing the evolution of machine learning applications in EELS analysis from 2010 onwards. Initial methods focused on principal component analysis (PCA) and basic noise reduction. Subsequently, clustering algorithms and matrix factorisation methods enabled unsupervised chemical mapping. Since 2020, the introduction of deep learning approaches such as convolutional neural networks (CNNs), autoencoders, and 208 List of Figures generative models has enabled more sophisticated analysis, including automated oxidation state classification and cross-technique data generalisation.125,131–139 .......................................... 27 Figure 2.13: Illustration of a tortuous path of length 𝛥𝑙 through a porous microstructure of thickness 𝛥𝑥, where the shortest tortuous path is used to calculate the tortuosity of the sample. (a) Geometric perspective without constrictions, (b) Geometric perspective with constriction, where lighter pink region represent less weightage compared to darker pink regions, which indicate more constricted areas ........................................................................................................ 29 Figure 2.14: Flowchart illustrating methods for measuring tortuosity: Geometric approaches using 3D imaging for direct analysis, physics-based methods for employing numerical simulations and electrochemical techniques like diffusion and symmetrical cell tests. Feedback loops and statistical analyses link microstructure to macroscopic transport properties for material optimisation. .............................................................................................. 30 Figure 3.1: Schematic representation of the electron beam interaction volume in a specimen, showing the various signals generated and their respective interaction depths160 .................. 35 Figure 3.2: The two lens condenser system. The spot of size s1 at the gun crossover (G) is demagnified to s2 by the first condenser lens C1. The second condenser lens C2 is used to focus the beam. (A), (B) and (C) show underfocused, focused and overfocused beams respectively. The convergence angle 𝛼 is controlled by the condenser aperture. ............................................ 38 Figure 3.3: Schematic of objective and intermediate lens operation in TEM. (a): When the intermediate lens focuses on the image plane of the objective lens, a magnified real-space image is formed. (b): When refocused on the back focal plane, a diffraction pattern is projected. The position of apertures and the lens configuration determine whether the output is an image or a diffraction pattern. ...................................................................................................................... 40 Figure 3.4: (a) Schematic of STEM detection geometry showing typical detector positions for EDX, EELS, and HAADF signal collection and (b) Angular ranges for scattered electrons captured by BF, ADF, and HAADF detectors, enabling complementary contrast based on scattering angle and atomic number.170 ........................................................................................... 43 Figure 3.5: Sampling regimes in STEM: (a) exact sampling, (b) undersampling, and (c) oversampling, based on the relationship between probe size (𝑑eff) and pixel size ................... 44 List of Figures 209 Figure 3.6: A schematic representation of the allowed electronic transitions from occupied atomic orbitals (core shells) to unoccupied states, as a function of increasing energy. The K, L, M, N, and O shells correspond to principal quantum numbers (n = 1, 2, 3, 4, 5), with sublevels split by orbital angular momentum (l) and spin-orbit coupling (j). Transitions labelled K, L₂,₃, M₂,₃, etc., represent characteristic core-loss edges observable in EELS, and are used for elemental identification and oxidation state analysis. The direction of transitions follows dipole selection rules (Δl = ±1).159 ..................................................................................................... 45 Figure 3.7: Probability of multiple inelastic scattering events against sample thickness and their corresponding 𝑡/𝜆 in vertical dashed lines. Each curve represents the probability of electrons undergoing zero (P0), single (P1), double (P2), or triple (P3) scattering events of NMC811 ............................................................................................................................................... 50 Figure 3.8: A converged electron probe is rastered across the specimen in STEM mode. At each probe position, the transmitted electrons are collected using a post-column EELS spectrometer, which records an energy-loss spectrum. Simultaneously, high-angle scattered electrons can be detected using a dark-field (DF) or HAADF detector. The resulting data forms a three- dimensional spectrum image (x, y, ΔE), where each pixel contains a full EELS spectrum. This data cube enables spatially resolved chemical and electronic structure analysis. 171 ............... 51 Figure 3.9: Beam-damage assessment for EELS acquisition conditions in NMC811. High- resolution TEM images and corresponding FFTs acquired before and after electron-beam exposure at two dose levels. Top: High-dose condition (500 pA, 1 s; ~3.12×109 e-) shows beam- induced damage with local amorphisation and altered FFT patterns. Bottom: Lower-dose condition (150 pA, 0.5 s; ~4.68×108 e-) shows no detectable structural change with preserved lattice fringes and unchanged FFT spots. ....................................................................................... 53 Figure 3.10: Schematic of modern dual-beam columns FIB systems. The electron beam and an ion beam allow for visualisation of the sample and deposition/milling in the region of interest. Figure from Carl Zeiss.174 .................................................................................................................. 55 Figure 3.11: Sequential SEM (orange border) and FIB (purple border) images illustrating the in-situ lift-out process for TEM lamella preparation. ................................................................... 56 Figure 3.12: Optical image of (a) a Cu half-grid from Agar Scientific showing multiple mounting positions (labelled A-C), and (b) close-up view highlighting finger B for 210 List of Figures conventional side-mounted lamella and finger C for the modified M-slot mounting. Red boxes indicate the regions where lamellae are typically attached using Pt deposition176................... 58 Figure 3.13: SEM (orange border) and FIB (purple border) images showing the modified lamella mounting and thinning strategy for improved mechanical stability and ultrathin TEM sample preparation ............................................................................................................................ 58 Figure 3.14: Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) workflow for 3D tomography of cycled NMC811 cathodes. (a) Schematic representation of the FIB-SEM slice- and-view process, showing the geometry of slice removal relative to the ion and electron beam axes. The protective platinum layer and fiducial marker are also indicated. Figure adapted from Liu et al.177 (b) SEM image of the sample following initial trench milling using a 30 nA gallium ion beam, showing redeposition around the milled volume. (c) Further polishing with a 1.5 nA ion beam reveals the detailed microstructure; vertical streaks (curtaining artefacts) are visible due to milling inhomogeneity. (d) Final high-resolution cross-sectional image used for segmentation, illustrating the distinct phases of the electrode microstructure. Scale bars in (b-d) represent 5 𝜇𝑚. .......................................................................................................................... 60 Figure 3.15 Examples of artefacts during FIB-SEM acquisition of NMC811: (a) Curtaining artefacts caused by ion channelling during milling, resulting in vertical striations; (b) Shine- through artefact, where material from deeper layers is imaged due to electrons escaping through an open pore; and (c) Sample charging visible as bright regions. Scale bar represents 2 μm. .................................................................................................................................................... 61 Figure 3.16 Comparison of curtain artifact removal techniques in SC2 FIB-SEM images. (a) shows the original image with visible curtain artifacts. The vertical detail coefficients isolated through the filtering process and the resulting image after applying the filter are as follows: (b) and (c) for wavelet-only filtering, (d) and (e) for FFT-only filtering, and (f) and (g) for the combined wavelet-FFT approach. Scalebar represents 2𝜇m. ....................................................... 62 Figure 3.17: (a) Schematic representation of the FIB-SEM tomography setup with sample stage tilted to 52°, positioning the ion beam perpendicular to the sample surface for uniform milling (b-d) origin of the shine-through artefact (STA), where electrons escape through large open pores, imaging deeper layers across multiple slices. In (c) slice n, the STA-affected region is highlighted; in (d), slice n+10 (approximately 500 nm deeper), (e) angled streaks in the Y-Z List of Figures 211 projection that arise because the artefact remains visible in nearly the same location across multiple slices. Figure produced by Morzy29 ................................................................................. 63 Figure 4.1: Schematic illustration of the calendering process and its microstructural effects on polycrystalline (PC) and single-crystal (SC) NMC811 particles. Calendering compresses the electrode using opposing rollers, applying a force (FN) that enhances interparticle contact. For PC particles (top row), the applied pressure can induce intergranular cracking due to the presence of grain boundaries. In contrast, SC particles (bottom row), lacking internal boundaries, undergo deformation while largely maintaining structural integrity. The light blue outlines represent the carbon-binder domain (CBD), which surrounds the active material particles and facilitates electronic conductivity and structural cohesion. ................................. 68 Figure 4.2: Schematic of Li-ion and electron transport in high-porosity (left) and low-porosity calendered (right) electrodes. Grey spheres: active material; black circles: carbon additives; purple regions: electrolyte-filled pores; green dashed lines: electronic percolation pathways; red regions: contact resistance (𝑅𝑐𝑜𝑛𝑡𝑎𝑐𝑡). Adapted from Kim et al.204 ..................................... 69 Figure 4.3: Schematic overview of the experimental workflow for investigating the effects of calendering on NMC811 cathode microstructure and transport properties. (1) Electrodes are calendered to different densities using heated rollers. (2) Samples are electrochemically cycled to simulate operational conditions. (3) FIB-SEM tomography is performed to acquire high- resolution 3D datasets. (4) The tomograms are segmented into distinct phases—active material, binder (CBD), and pore. (5) Tortuosity and transport pathways are quantified via graph-based analysis to evaluate directional anisotropy and the impact of microstructural changes. ............................................................................................................................................... 73 Figure 4.4: Example of 3D-UNet based segmentation of FIB-SEM tomography data based on secondary electron detector images. The SEM images were reconstructed in the Y-Z plane (perpendicular to the FIB slicing plane, current collector is towards the bottom of the image, separator side towards the top). Overlay of the segmented regions of (a) shine-through artefacts (STA), (b) carbon binder matrix, (c) pore, (d) active material. ..................................... 77 Figure 4.5: (a) The reconstructed volume of calendered single crystal electrode (SC-25) from FIB-SEM imaging. The blue arrow corresponds to the through-plane direction, which is the direction of calendering, perpendicular to the current collector, red arrow corresponds to in- 212 List of Figures plane direction, parallel to the current collector and green arrow represents the direction of serial sectioning during FIB-SEM acquisition. (b) Schematic comparison of sparse graph path (yellow), dense graph path (red) and Euclidean path (purple). The sparse graph path simplifies connectivity with fewer nodes and edges, offering computational efficiency, while the dense graph path captures all possible connections for detailed structural analysis. The Euclidean path represents the straight-line distance between the start and end points. ........ 78 Figure 4.6: Active material in grey with overlaid pore network visualisation (white nodes with multi-coloured edges representing the difference in lengths), demonstrating the complex transport pathways through the structure of PC in (a), (b) and (d), and SC in (b). ................. 79 Figure 4.7: Comparison of weighted (left) and non-weighted (right) graphs illustrating the shortest path calculation using Dijkstra’s algorithm. In the weighted graph, edge weights influence the shortest path selection, leading to a route that minimises the total weight (red edges). The non-weighted graph treats all edges as equal, resulting in a shortest path based solely on the number edges. ............................................................................................................. 81 Figure 4.8: Illustration of the 3D volume analysis approach for TauFactor tortuosity calculations. (a-b) Representative 2D binary images (single slices from the 3D FIB-SEM tomography datasets) showing segmented pore phase (white) for polycrystalline (PC) and single crystal (SC) electrodes. (c) 3D subvolume approach: The reconstructed 3D volume is divided into four 5 × 5 𝜇𝑚 subregions (numbered 1-4, shown in different colors), each analysed independently through TauFactor to calculate tortuosity across the entire 3D stack. (d) 3D full-volume approach: A larger 10 × 10 𝜇𝑚 depth domain is analysed as a single 3D dataset. The red dashed squares in panels (a-b) correspond to the 2D projection of the subvolume regions shown in 3D in panel (c), while the blue dashed squares correspond to the full-volume region in panel (d). ....................................................................................................... 83 Figure 4.9: (a) General transmission-line model (TLM) equivalent circuit for a porous electrode, representing charge transport in the solid (grey region) and electrolyte (blue region) phases. The solid phase is characterised by electronic resistances, rEl , while the electrolyte phase includes ionic resistances, rion . The charge transfer process at the solid-liquid interface is modelled by surface impedance elements, zs , which account for faradaic or capacitive reactions. (b) Simplified transmission-line model for porous electrodes under blocking List of Figures 213 conditions and with negligible electronic resistance (rEl ≪ rion), including ionic resistances, rion, and constant-phase elements, qs for capacitive behaviour. (c) Schematic representation of the porous electrode microstructure, highlighting an ionic pathway within the electrolyte phase through a crack in the structure, representative of the transport pathways modelled in the TLM. Adapted from refs157,221 ..................................................................................................... 85 Figure 4.10: Cross-section images of uncalendered, calendered to 35% and 25% porosities of poly and single crystal NMC811 prior to any electrochemical cyclings. Scalebar for the top and bottow rows are 2 𝜇𝑚 and 5 𝜇𝑚 respectively. ................................................................................ 89 Figure 4.11: SEM surface images of single-crystal (top row) and polycrystalline (bottom row) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The red arrows indicate cracks that can be seen from the surface of calendered PCs. The scalebar is 10 𝜇𝑚. .................................................................. 90 Figure 4.12: Cross-sectional SEM images of single-crystal (SC) and polycrystalline (PC) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The top row shows SC samples, where the structure remains largely intact across all porosity levels. The bottom row depicts PC samples, illustrating progressive intergranular cracking and structural degradation as porosity decreases. Red arrows in the SC samples highlight intragranular cracks, while the PC samples demonstrate severe compaction and cracking, particularly for calendered samples, indicative of higher mechanical stress and calendering-induced damage. The scalebar is 2 𝜇𝑚. ............ 91 Figure 4.13: Magnified cross-sectional SEM images of single-crystal (SC) and polycrystalline (PC) NMC811 electrodes after 300 cycles at a 1C rate, prepared with varying porosities: uncalendered, 35%, and 25% porosity. The top row shows SC samples with intragranular cracks. The bottom row illustrates PC samples, highlighting significant intergranular cracking that increases in density and severity with decreasing porosity. The scalebar for top and bottom rows are 1 𝜇𝑚 and 2 𝜇𝑚 respectively. ............................................................................... 92 Figure 4.14: Cross-sectional SEM images of a polycrystalline (PC) NMC811 electrode after 300 cycles at 1C rate acquired using slice-and-view technique, with each image representing the 10th slice along the z-axis at 600 nm intervals. Scalebar represents 2 𝜇𝑚. ................................. 93 214 List of Figures Figure 4.15: Comparison of tortuosity values derived from dense and sparse graphs for (a) Active Material and (b) Pore in the throat-weighted analysis along direction of calendering. Error bars represent the standard deviation of the pathway tortuosity distribution reported by Dragonfly for the selected input-output boxes. ....................................................................... 94 Figure 4.16: Comparison of tortuosity values derived from throat weighted and non-throat weighted for active material and pore along the direction of calendering. Error bars represent the standard deviation of the pathway tortuosity distribution reported by Dragonfly for the selected input-output boxes. ............................................................................................................ 95 Figure 4.17: Comparative FIB-SEM images of single-crystal (top row) and polycrystalline (bottom row) electrodes at 25% porosity. Scale bar: 5 μm. .......................................................... 97 Figure 4.18: Comparison of tortuosity values calculated using the full-volume (solid triangles) and sub-volume (open triangles) approaches for single-crystal (SC) and polycrystalline (PC) samples at varying porosities (43%, 35%, and 25%) The error bar represents the standard deviation between the four sub-volumes. ...................................................................................... 98 Figure 4.19: (a) Nyquist plot of symmetric cell assembled with polycrystalline (PC) and single crystal (SC) NMC811 under blocking conditions. (b) Calculated MacMullin numbers using the ionic resistance obtained from the Nyquist plot. .................................................................... 99 Figure 4.20: The tortuosity-porosity relationship of (a) PC with Bruggeman fitting, (b) PC with Bruggemen fitting and pre-exponential factor and (c) SC. ........................................................ 100 Figure 4.21: Tortuosity measurements for porous network across different sample types and measurement methods. (a-c) Method-specific comparisons showing through-plane (solid lines, dark colours) and in-plane (dashed lines, light colours) tortuosity for (a) Dragonfly, (b) Taufactor, and (c) EIS measurements across SC (uncompressed, 35%, 25% compressed) and PC (uncompressed, 35%, 25% compressed) samples. (d-e) Direction-specific comparisons showing (d) through-plane and (e) in-plane tortuosity for all three methods (Dragonfly: green circles, Taufactor: orange triangles, EIS: brown squares). Error bars represent standard deviation. The vertical dotted line separates SC and PC sample types. Note: EIS measurements were only performed in the through-plane direction. ........................................................................................................... 105 List of Figures 215 Figure 4.22: Tortuosity anisotropy in NMC811 electrodes for pore and active-material phases. Tortuosity anisotropy is defined as 𝜏through-plane/𝜏in-plane . (a) Pore-phase anisotropy and (b) active-material-phase anisotropy, extracted using Dragonfly graph-based analysis and TauFactor for each sample (SC-UC, SC-35, SC-25, PC-UC, PC-35, PC-25). The dashed horizontal line at unity indicates isotropic transport (𝜏through = 𝜏in), while values above (below) 1 indicate higher tortuosity in the through-plane (in-plane) direction. The dotted vertical line separates single-crystal (SC) and polycrystalline (PC) electrodes. ......... 107 Figure 4.23: Schematic illustration of ion and electron transport pathways in single-crystal (SC) and polycrystalline (PC) NMC811 electrodes at different porosities. Top row: PC electrodes showing increased intergranular cracking and tortuous pathways with calendering. Bottom row: SC electrodes demonstrating enhanced alignment and reduced tortuosity with calendering, promoting more direct ionic and electronic transport pathways. Pink arrows indicate ion (Li+) movement, and yellow arrows show electron transport through the carbon black network (blue) through grain boundaries or from one particle to another .................. 112 Figure 5.1: Schematic representation of all variable current protocols used, the C-rate applied per interval and how much time it would take to complete 0.5C equivalent. (a) is the commonly used CCCV, (b) the LowEnd, (c) the LowStart. This is only representative for the charge sequence. Protocols set by Dr Alexander Dimitrijevic. .................................................. 118 Figure 5.2: EELS spectrum of NMC811 showing the characteristic absorption edges of oxygen (O-K1, O-K2), manganese (Mn-L2, Mn-L3), cobalt (Co-L2, Co-L3,), and nickel (Ni-L2, Ni-L3), used to analyse oxidation states .............................................................................................................. 121 Figure 5.3: (a) Monochromated EELS spectrum of Ni2+ showing clear multiplet splitting in the Ni-L3 edge, with distinct peaks corresponding to Ni-L3,high and Ni-L3,low, separated by approximately 1.7 eV. The high energy resolution (FWHM ~0.25 eV) enables the resolution of fine electronic structure features. Data provided by colleague, Mr Wei Huang (b) Non- monochromated EELS spectrum of Ni, acquired with 0.3 eV/channel dispersion, where the Ni-L3 edge appears as a single broadened peak with no resolvable multiplet splitting ........ 122 Figure 5.4: (a) HAADF STEM image of NMC811 cycled under the LowStart protocol with a selected spectrum image region highlighted in the green box. The corresponding maps of the spectrum image are shown in (b) for t/λ and (c) for thickness in nm. ..................................... 125 216 List of Figures Figure 5.5: Standard background subtraction using PowerLaw functions in HyperSpy. (a) Oxygen K-edge and (b) Nickel L-edge spectra with poor baseline fitting (grey dashed line) ............................................................................................................................................................ 126 Figure 5.6: Background subtraction methodology for EELS analysis of NMC811 cathode materials. Panels show (a) Ni L-edge, (b) Mn L-edge, (c) O K-edge and (d) Co L-edge with original data (red), fitted background (blue dashed) constrained by pre-edge (light yellow) and post-edge (light orange) regions, and resulting background-removed signals (black). Coloured regions highlight the relevant L3 and L2 edges for transition metals and pre- peak/main peak structures for oxygen .......................................................................................... 127 Figure 5.7: (a) Single Gaussian function defined by its peak height ℎ, centre position 𝜇, and standard deviation 𝜎, which controls the width of the curve. (b) Double Gaussian function composed of two individual Gaussian components: one with parameters ℎ1, 𝜇1 , 𝜎1 (green) and the other with ℎ2, 𝜇2 , 𝜎2 (blue). The sum of both components gives the total fitted peak 𝐺𝑡𝑤𝑜 (red) .......................................................................................................................................... 129 Figure 5.8: Definition of energy range and peak range used in Gaussian fitting of the Ni L3 and L2 edges. The energy range (blue arrows) represents the full window selected to isolate the spectral feature, while the peak range (black arrows) defines a narrower fitting window (±2 eV) centred around the peak maximum to focus the fitting and avoid interference from neighbouring features. .................................................................................................................... 130 Figure 5.9: Multi-Gaussian fitting of EELS spectra for key edges in NMC811. Double Gaussian fitting for (a) O-K1 pre-edge region, (c) Mn-L3, (d) Mn-L2, (e) Co-L3, (f) Co-L2, (g) Ni-L3, (h) Ni- L2 edges, and triple Gaussian fitting of (b) O-K2 main edge. Grey circles represent raw EELS data; black line indicate individual Gaussian components (g1, g2, g3), and their summed fit in their respective colours. Intensity values have been normalised to 1. ..................................... 133 Figure 5.10: Background-subtracted EELS spectra of LowEnd, LowStart, and pristine NMC811 samples at both surface (lighter shade) and bulk (darker shade) regions. Spectra for (a) O K-edge, (b) Ni L-edge, (c) Mn L-edge, and (d) Co L-edge ................................................ 134 List of Figures 217 Figure 5.11: Comparative depth profiles of L3/L2 area ratios and oxygen ΔE parameter across different cycling protocols. Fluctuations in Mn and Co trends are attributed to low signal-to- noise. Shaded halos represent error bars based on the dispersion channel during EELS acquisition. ........................................................................................................................................ 138 Figure 5.12: Depth profiles of L3/L2 area ratios for Ni, Co, and Mn (left axis) and oxygen ΔE parameter (right axis) for NMC811 samples cycled under LowStart, LowEnd, and pristine conditions. Fluctuations in Mn and Co trends are attributed to low signal-to-noise. Shaded halos represent error bars (±0.3 eV) based on the dispersion channel during EELS acquisition ............................................................................................................................................................ 139 Figure 5.13: Spatial evolution of EELS peak centre positions from the particle surface towards the bulk across a LowStart-cycled NMC811 particle. (a-b) Ni L3 and L2 edges, (c-d) Co L3 and L2 edges, (e-f) Mn L3 and L2 edges, and (g-h) O K pre-peak and main peak positions. Error bars represent the conservative uncertainty associated with the spectral dispersion (0.3 eV per channel) and fit stability, as described in the text. The maps visualise the systematic near-surface shift relative to the particle core that underpins the protocol comparison. ...... 141 Figure 5.14: Atomic-resolution imaging of structural phases in LowEnd NMC811. (a) HAADF- STEM image showing the RSL-spinel-layered transition. (b-d) High-magnification images overlaid with structural models of each phase. Atom colours: yellow = Li, blue = Transition metal, red = O.................................................................................................................................... 143 Figure 5.15: Cross sectional HAADF-STEM overview images of NMC811 particles after cycling under (a) LowStart, (b) LowEnd, and (c) pristine conditions. Note the extensive cracking in LowEnd versus the more intact morphology of LowStart .................................... 144 Figure 5.16: HAADF-STEM analysis of crack-induced degradation for LowStart sample. Line profiles show a transition from rock-salt (~2.54 Å) near the crack to layered structure (~4.96 Å) deeper into the particle. ............................................................................................................. 145 Figure 6.1: Histograms showing the distribution of L3/L2 ratios measured from reference compounds: NiO (Ni2+), LiNiO2 (Ni3+), and delithiated LiNiO2 (Ni4+). Note the significant overlap between Ni3+ and Ni4+, and partial overlap between Ni2+ and Ni3+. ............................ 152 218 List of Figures Figure 6.2: Classification maps based on L3/L2 ratios for (a) unbinned LowStart sample, (b) binned LowStart sample, (c) unbinned LowEnd sample, and (d) binned LowEnd sample. Blue regions represent Ni2+ (L3/L2 > 3.41), purple indicates mixed valence (3.31-3.41), and orange shows Ni3+ (L3/L2 < 3.31). The x-axis represents distance from the particle surface in nanometers. ....................................................................................................................................... 153 Figure 6.3: Comparison of reference spectra for Ni oxidation states. (a) Mean reference spectra for Ni2+ (blue) and Ni3+ (orange) showing distinct differences in L3 and L2 peak shapes. (b) Ten Ni3+ reference spectra showing consistency in L3 peak shape despite some noise in the L2 region. (c) Ten Ni2+ reference spectra showing sharper, more symmetric L3 peaks compared to Ni3+. ..................................................................................................................................................... 155 Figure 6.4: Histograms showing the distribution of L3 peak areas for reference sample of Ni2+ (blue) and Ni3+ (orange). .................................................................................................................. 156 Figure 6.5: Histograms showing the distribution of L3 peak areas for (a) LowStart, (b) LowEnd, and (c) pristine samples. The dashed red lines indicate the boundaries between Ni2+ (2.5-2.9), mixed (2.9-3.4), and Ni3+ (3.4-4.5) states. Insets show the pixel distribution by oxidation state across the sample depth. ................................................................................................................. 157 Figure 6.6: Representative reference spectra for (a) Ni3+ and (b) Ni2+ used in CNN training, including spectral augmentations to simulate energy drift. Each panel shows ten individual spectra (solid lines), with the dotted line indicating the mean spectrum for each oxidation state. Augmentation was applied to improve the model's robustness to experimental variability and ensure generalisability across different acquisition conditions. ..................... 159 Figure 6.7: Detailed architecture of the CNN model used for nickel oxidation state classification. The network features multiple convolutional layers with varying kernel sizes, an attention mechanism, and a global average pooling branch working in parallel to extract diverse spectral features. Conv1d(x, y, k) represents a 1D convolutional layer with x input channels, y output channels, and k kernel size. FC(x, y) represents a fully connected layer with x input features and y output features. Architecture designed by Kelvin Chan in collaboration with May Ching Lai. ........................................................................................................................ 160 List of Figures 219 Figure 6.8: Confusion matrix for CNN classification of nickel oxidation states using EELS spectra. The matrix shows results for 4,800 validation spectra. Diagonal cells represent correct classifications, True Positives (TP): 2,402 Ni2+ spectra and 2,339 Ni3+ spectra were correctly identified. Off-diagonal cells represent misclassifications: 59 Ni3+ spectra were incorrectly labelled as Ni2+, as False Positives (FP) for Ni2+ or False Negatives (FN) for Ni3+, while no Ni2+ spectra were misclassified. Note that reference labels are assumed ground truth and have not been independently verified by complementary techniques such as XAS or XPS. ................ 165 Figure 6.9: CNN-classified oxidation state maps for (a,c,e) discrete classification and (b,d,f) probability visualisation of LowStart, LowEnd, and pristine NMC811 samples, respectively. Blue regions represent Ni2+-rich areas, while orange indicates Ni3+-dominated regions. In the probability maps, the color gradient indicates the model's confidence in classification, with white regions representing approximately equal probability between Ni2+ and Ni3+, suggesting mixed valence states. The x-axis represents distance from the particle surface in nanometers. ............................................................................................................................................................ 168 Figure 6.10: Comparative spectral analysis of Ni oxidation states in NMC811 samples. Left column (a,d,g): Average spectra of pixels classified as Ni2+ (blue) compared to Ni2+ reference (red). Middle column (b,e,h): Average spectra of pixels classified as Ni3+ (orange) compared to Ni3+ reference (red). Right column (c,f,i): Comparison of mean spectra for all pixels classified as Ni2+ (blue) versus Ni3+ (orange) within each sample. Top row: LowStart sample; Middle row: LowEnd sample; Bottom row: Pristine sample. All test spectra were energy-aligned with reference spectra to facilitate direct comparison of peak shapes. ............................................. 170 220 List of Tables List of Tables Table 2.1: Summary of key battery performance metrics .............................................................. 8 Table 3.1: Characteristics of the principal electron sources at an acceleration voltage of 100 kV.13 ...................................................................................................................................................... 37 Table 3.2: Microscope parameters and material constants used for multiple scattering analysis, based on a beam energy (𝐸0) of 300 kV, a collection semi-angle of 17 mrad, and a relativistic factor (F) of 0.514. 𝑍𝑒𝑓𝑓 is the effective atomic number, 𝐸𝑚 is the characteristic energy loss per inelastic scattering event, and 𝜆 is the inelastic mean free path. For NMC811, 𝜆 values are compared across three structural phases: layered Li(Ni0.8Mn0.1Co0.1)O2, spinel Li(Ni0.5Mn1.5)O4, and rock-salt Ni2O. Additional reference materials, including hybrid perovskite, Cs0.1[CH(NH2)2]0.9Pb(I0.955Br0.045)3, along with WO3, SiO2, and SnO2 are included for comparison to NMC811. ......................................................................................................................................... 47 Table 3.3: The probabilities for zero (P0), single (P1), double (P2), and triple (P3) scattering events at selected sample thicknesses and their corresponding 𝑡/𝜆 values, at 300 keV and 17 mrad ..................................................................................................................................................... 49 Table 4.1: Compilation of electrochemical test results carried out by Dr Kumar Raju for SCs and PCs at varying porosity levels. UC = Uncalendered 114 ....................................................... 103 Table 4.2: Comparison of tortuosity calculation methods for NMC811 electrodes ............... 110 Table 5.1: Start and end discharge capacity (normalised) for graphite/NMC811 full cells cycled at ambient temperature under CCCV, LowStart and LowEnd protocols. Electrochemical data provided by Dr Alexander Dimitrijevic; full electrochemical characterisation available in Dr Dimitrijevic's thesis.251 ..................................................................................................................... 120 Table 5.2: Optimised background fitting parameters for EELS analysis of NMC811 components ....................................................................................................................................... 127 List of Tables 221 Table 5.3: Spectral fitting constraints for core-loss edges in NMC811. The energy range defines the total span considered for fitting, and the peak range indicates the ±eV window centred around the detected maximum for refining the fit. ** denotes double gaussian and *** denotes triple gaussian used for fitting. ...................................................................................................... 131 Table 5.4: Reference L3/L2 ratios and peak positions for key oxidation states. Ni references acquired in-house; Mn from SuperSTEM; Co and additional Mn values adapted from literature,263 denoted as * for Mn to distinguish between SuperSTEM and reference values. ............................................................................................................................................................ 136 Table 5.5: Summary of L3/L2 ratios, ΔE, and peak centre positions for pristine, LowStart, and LowEnd samples. Mn and Co values fluctuate due to low signal. ΔE indicates O-K pre/main peak separation; degradation zone depth is estimated from combined redox and peak shift profiles. .............................................................................................................................................. 142 Table 6.1: Summary of CNN training and validation dataset composition. ........................ 160 Table 6.2: Glossary of CNN Terms Used in the Architecture276–279 ............................................ 162 Table 6.3: Classification performance metrics for CNN-based Ni oxidation state determination. ............................................................................................................................................................ 166 Table 6.4: Test dataset characteristics and CNN classification results. ................................. 167 Table 6.5: Comparison of different approaches for nickel oxidation state classification using EELS. .................................................................................................................................................. 176 222 Appendix Appendix Figure A1: Elemental edge peak positions as a function of distance from the particle surface in a cycled NMC811 sample (LowEnd protocol). Peak centre positions were extracted from background-subtracted EELS spectra for each transition metal and oxygen edge. A gradual redshift is observed in the Ni L₃, Ni L₂, and O pre-peak edges, indicating progressive reduction and oxygen loss at the surface, consistent with the formation of a surface reconstruction layer. Co and Mn edges show minimal variation, suggesting more stable oxidation states across the profile Appendix 223 Figure A2: Elemental edge peak centre positions extracted from background-subtracted EELS spectra across the surface-to-bulk direction of a pristine NMC811 particle. The Ni L₃ and L₂ edges (top row) exhibit minimal variation with distance, indicating uniform oxidation state throughout the particle. Co and Mn edges (middle rows) also show stable peak positions, suggesting minimal redox activity or electronic structure distortion. The O K-edge pre-peak and main peak positions (bottom row) remain consistent, further confirming the structural and chemical homogeneity of the pristine material. 224 Appendix Figure A3: L3/L2 area ratios and peak centre positions of Ni2+ and Ni3+ from NiO and LiNiO reference sample respectively Ni 2+ Ni 2+ Ni 3+ Ni 3+ Appendix 225 Figure A4: L3/L2 area ratios and peak centre positions of Mn2+ and Mn3+. Data provided by Dr Demie Kepaptsoglou at SuperSTEM Mn 2+ Mn 2+ Mn 3+ Mn 3+ 226 Appendix Figure A5: L3/L2 area ratios and peak centre positions of Mn4+ from NiO and LiNiO reference sample respectively. Data provided by Dr Demie Kepaptsoglou at SuperSTEM Mn 4+ Mn 4+ Appendix 227 Figure A6: CNN classification results for Ni oxidation states based on EELS spectra for (a) LowStart-cycled, (b) LowEnd-cycled, and (c) pristine NMC811 samples. Each subpanel displays the predicted class labels alongside probability maps for Ni2+ and Ni3+, using (d) input spectra from 840 - 880 eV to cover both the Ni L3 and L2 edges. Across all samples, the model consistently assigns high-confidence predictions for Ni3+, even in regions expected to show surface reduction (Ni2+), such as in the cycled samples. This systematic misclassification highlights the CNN’s inability to accurately learn and distinguish the combined L3 and L2 edge shapes, which exhibit subtle but critical variations in degraded zones. The result reflects a key limitation in the current model architecture and training set, suggesting the need for more representative data and improved spectral feature learning strategies.