We evaluate the performance of SEM-EDS-based mineral identification and quantification by comparison to XRD-derived quantitative data, and by checking that grains identified in SEM-EDS match with grain outlines and characteristic mineral morphologies in backscatter electron (BSE) images. For these comparisons, 16 fine-grained rock samples were selected from the shale interbeds of the second member of the Ediacaran Doushantuo Formation at two different sites in the Yangtze Gorges Area, South China (Han et al., 2022): 1) the Sixi Section (30°45′25.8″ N, 110°55′51.0″ E) and; 2) the Qinglinkou Section (30°48′0.2″ N, 110°55′18.5″ E), which together include one black shale, 10 dolomitic shales and five shaly dolostones (Supplementary Table S1). A complementary dataset was obtained from Enorama Shale (Cryogenian Umbertana Group), Brachina Formation (Ediacaran Wilpena Group) and Bunyeroo Formation (Ediacaran Wilpena Group), which crop out within Brachina Gorge, central Flinders Ranges (South Australia; Rafiei et al., 2020). These locations include six finely laminated siliciclastic mudstones to siltstones (Rafiei et al., 2020) and span the Global Stratotype Section and Point (GSSP) for the Ediacaran Period. These two sets of samples are both sedimentary archives close to the boundary between Precambrian and Cambrian, documenting major environmental changes in Earth history. This combined sample set covers the majority of fine-grain lithologies found at the Earth surface, and provides good examples to understand detrital versus authigenic mineral assemblages (Rafiei et al., 2020).

Sample blocks with a 1–2 cm² surface were cut perpendicular to the bedding plane and prepared as 30 mm diameter resin mounts with the cross section exposed. The resin mounts were polished by standard mechanical polishing techniques and subsequently ion milled using a Hitachi IM4000 Argon Ion Mill (15 min, 12° beam incident angle, 5 kV accelerating voltage and continuous sample rotation) to remove remaining surface damage. The specimens were then carbon coated for subsequent SEM imaging and mineral mapping.

The SEM imaging and SEM-EDS mineral mapping were carried out using a FEI Teneo LoVac field emission instrument equipped with dual Bruker XFlash 6/30 EDS detectors. Whole-of-sample (overview) BSE image tile sets (12 mm working distance; 15 kV accelerating voltage) were obtained and stitched using the Maps Mineralogy software (version 3.1). These overview BSE images were used to identify representative sample areas for optimal ROI placement. High-resolution BSE images (100 nm pixel resolution) and EDS spectra (1–2 µm step size/pixel resolution, 8 ms acquisition time, 3,000–6,000 counts per pixel) were subsequently collected for each ROI (Figure 1). One to two ROIs were analyzed per sample (20 ROIs over 16 samples), with ROI size ranging between 0.08 and 9.19 mm² (Table 2; see Section 3.3). EDS spectra were classified using the Nanomin software (version 3.18), where mineral identification is achieved by comparison of individual EDS spectra to reference spectra collected on known mineral standards (Abbott et al., 2019; Frank et al., 2020; Rafiei et al., 2020). A full list of minerals included in the “mineral recipe” used for classification and the measurement parameters are given in Supplementary Table S2; the mineral list was informed by independent XRD-based mineral characterization of the samples. Mineral concentrations (reported as wt%) in the mapped areas are determined by converting the mineral area % to wt% using published densities for all identified sample constituents.

The Nanomin algorithm classifies individually measured EDS spectra by computing an error score relative to a subset of mineral reference spectra stored in a reference library (organized in user-defined “mineral recipes”), with the resulting classification being the mineral that incurs the lowest error score. Importantly, the algorithm can deconvolve mixed EDS spectra and assign up to three minerals per analyzed single spot (producing “mixels”), explicitly recognizing and resolving complex mineral mixtures that occur in the electron-beam excitation volume of a typical fine-grained material. Together with the implementation of “mixels”, the implementation of an error minimization-based classification with a strict goodness of fit cutoff (90% by default) has the potential to significantly improve mineral identification and quantification compared to previous generations of SEM-EDS mineral mapping tools, whilst requiring substantially less analyst input.

We evaluate the quality of mineralogy data from SEM-EDS mineral mapping by comparison to quantitative XRD results (see Section 3.2). To this end, we calculate the residual between mineral abundances obtained from SEM-EDS mineral mapping and XRD analysis [wt% (Residual) = wt% (SEM-EDS mineral mapping) - wt% (XRD)]. To determine the impact of spectral deconvolution allowing multiple minerals to be distinguished in each analytical volume, we process each dataset using both: 1) the three minerals per pixel (“mixel”) and 2) the one mineral pixel mode, and subsequently compared the results for each sample (see Section 3.3). To explore optimal ROI selection (see Section 3.4), sub-ROIs of varying size (∼5 mm² and smaller) and geometry are placed on two samples that represent a homogenous lithology (SX22_1) and a heterogenous, laminated lithology (Qlk 6).

Subsamples for XRD analysis were cut next to the mineral mapping subsamples, crushed to powder in a Rock Labs RM100 Disc Mill and micronized (1–5 μm size range) with ethanol (Kleeberg et al., 2008) for 5 minutes in a McCrone mill with corundum beads. Samples were then oven dried and homogenized, before back loading into 23 mm stainless steel sample holders for analysis. XRD patterns spanning 3.5°–95° 2θ were collected using a Malvern Panalytical Aeris tabletop XRD instrument (Co radiation source with 40 kV generator voltage and 15 mA tube current, PIXcel3D-Medipix3 1×1 detector, 1/4-inch divergence slits) with 0.022° step size and total count time of 13 min. Diffraction patterns were interpreted using the Panalytical HighscorePlus software with the ICSD database for phase identification, phase quantification utilized Rietveld refinement as described in Richoz et al. (2017).

SEM-EDS mineral mapping and XRD analysis show that the Doushantuo Formation shale, dolomitic shales and dolostones mainly consist of dolomite (0–75.3 wt%), quartz (3.8–62.6 wt%), feldspar (0–33.1 wt%), clay minerals (4.3–34.4 wt%) and some accessory minerals (<0.5 wt%) such as rutile and zircon (Figures 1G,H; Figure 2A). Clay minerals are the main phases in the fine grain fraction of all samples. The dominant clay phases are smectite (saponite), chlorite (Mg-rich; clinochlore) and smectite-chlorite mixed-layer clay, often associated or intergrown with each other as described in more detail in Section 4.2. Brachina Gorge samples from Rafiei et al. (2020) are siliciclastic siltstones to mudstones. They have similar ranges of quartz and feldspar as the Doushantuo Formation samples, but with a higher clay mineral content (29–59 wt%) and only minor carbonate content (0–6 wt%) (Figure 2D). The Brachina Gorge samples contain a higher abundance of clastic material than the Doushantuo Formation samples. Their clay mineral contents are dominated by illite and chlorite with no smectite.

We evaluate the accuracy of Nanomin mineral identification using two approaches. First, we compare the mineral boundaries assigned by Nanomin to the mineral boundaries in co-located BSE images. Overall, this shows good agreement, with only a few unclassified pixels (Figures 1, 2; Table 2). The diagnostic grain boundaries and differences in contrast identifying the diagnostic crystal forms (e.g., dolomite, pyrite) evident in BSE images (Figures 1, 2, 9) closely align with the boundaries defined by SEM-EDS mineral mapping.

Secondly, we compare SEM-EDS- and XRD-based mineral quantification. This shows a strong linear correlation (see Section 3.2) and near-identical mineral ranges (Supplementary Tables S1,S3) except for minor minerals such as rutile and zircon, minerals that mostly occur at concentrations beneath the XRD detection limit (León-Reina et al., 2016).

Up to eleven major mineral phases were detected both by mineral mapping and XRD in the 16 Doushantuo Formation samples, equating to 220 XRD and SEM-EDS mineral mapping data points across the 20 ROIs (Figures 3A, Supplementary Table S1). Together with the 48 previously published Brachina Gorge data points from Rafiei et al. (2020), this comparison of SEM-EDS and quantitative XRD data points shows a strong linear correlation (R ² = 0.98) with the trendline falling very close to the ideal 1:1 line (slope = 0.97) and a y-intercept that is very close to 0 (0.25; Figure 3A). The strength of this relationship at the individual mineral level is also strong, but does show some variation. Minerals which occur at higher abundances or across a greater range (e.g., quartz, dolomite, feldspars, illite/muscovite) generally display a stronger correlation between XRD and SEM-EDS than low abundance/range minerals (e.g., pyrite, kaolinite) (Table 1; Figures 3B–D). In general, the slope and R ² of the trendlines are closer to 1.0 for higher abundance minerals. Minerals present at elevated abundances (>10 wt%) have R ² values larger than 0.85 and slopes between 0.8 and 1.2, while less abundant minerals comprising <10 wt% (mostly <5 wt%, except 6 wt% for hematite) show lower R ² values or scatter away from the theoretical 1:1 line (Figures 3B–D).

To evaluate the effectiveness of the multi-phase approach, which produce “mixels” to deal with mixed phases, we reprocessed the same dataset allowing only one mineral to be identified per pixel and compared it with the results presented in Section 3.2 (which featured mixels with up to three minerals per pixel). Our results show that an average of 28.6% of each ROI remains unclassified when processed using the single pixel mode, reflecting the abundance of mixed phases in the X-ray generating volume across major parts of these fine-grained samples. The strict error (goodness of fit) cut-off we use (90% spectral match) means that these mixed signals are not matched to any single mineral reference spectra. Consistent with this, most unclassified pixels are on the boundaries of minerals and within the fine-grain fraction (Figure 4). In contrast, where the data is processed permitting a maximum of three minerals per pixel “mixel” mode, the average unclassified area of each ROI is only 1.1% (Table 2), a reduction by more than one order of magnitude.

Surprisingly, the dataset processed using the single pixel mode still shows a strong correlation between SEM-EDS mineral mapping and XRD results (R ² = 0.94, Supplementary Table S3). This is because the conversion of mineral area% to mineral wt% involves the normalization of total mineral areal abundances to 100 wt%. While this step corrects to some degree for the high proportion of unclassified pixels, close inspection of our results shows that this normalization also masks systematic errors associated with the single pixel mode. Importantly, there is significantly more scatter away from the 1:1 line than in the three mineral “mixel” mode (compare Figure 3A versus Figure 5A). As a result, the mean residual between mineral mapping and XRD mineralogy for the one mineral per pixel mode is 2.5 times larger than in the “mixel” mode (residuals of 2.3% compared to 0.9% respectively). Furthermore, not all minerals are impacted to the same degree by this correction. The abundance of coarser grained minerals such as fine sand to silt size quartz, feldspar and dolomite is systematically overestimated in the one-pixel mode, showing trendlines with slopes larger than 1, whereas fine-grained components are systematically underestimated with slopes consistently lower than 1 (e.g., slope is 0.76 for smectite; Figure 5B, Table 3). This pattern reflects the tendency of coarse-grained material to produce “pure” or single-phase X-ray spectra, except at grain boundaries, whereas fine-grained materials are more consistently associated with other fine-grained mineral components and hence more likely to produce mixed-phase X-ray spectra that remain “unclassified” (unless the spectral deconvolution of the “mixel” mode is employed).

In summary, the large reduction in unclassified pixels and better match to bulk XRD shows that the automated spectral deconvolution of “mixel” mode accurately and effectively deals with mixed phases. This approach has the advantage of being a universally applicable, mathematical error minimization approach which does not require active user input or definition of bespoke mixed phases.

Our results demonstrate that SEM-EDS mineral mapping can accurately identify individual mineral grains and, in aggregate, produce quantitative mineralogical data consistent with that obtained by bulk sample XRD analysis. However, given that samples can be texturally and mineralogically heterogenous, appropriate ROI placement and geometry is key for obtaining reproducible bulk mineralogical data from SEM-EDS mineral mapping. We explore this here by comparing the impact of ROI placement and ROI size on bulk mineral results obtained for two very different samples: the first sample is relatively homogeneous (SX22_1; Figure 6), the other sample is laminated with major micro-scale textural and mineralogical variability (Qlk6; Figure 6). Both samples show good correlation between mineral mapping results and quantitative XRD results when using large ROIs (∼5 mm²) placed to ensure they are representative of the overall sample, with residuals consistently below 3 wt% (average <1 wt%).

“We divided larger ROIs of both samples” larger ROIs into 2 (∼2.5 mm²), 4 (∼1.3 mm²), 8 (∼0.6 mm²), 16 (∼0.3 mm²) and 32 (∼0.15 mm²) equal-sized sub-ROIs (Figure 6). Comparison of SEM-EDS-based mineralogical data from the smaller sub-ROIs with XRD results shows that the greatest differences in mineral abundances between the two methods occurs with SEM-EDS-based mineralogy data obtained from the smallest sub-ROIs (Figure 6). While the average and median difference in SEM-EDS-based mineral abundance across all sub-ROIs versus XRD data only increases slightly with decreased sub-ROI size, the frequency and magnitude of outliers increases substantially with decreasing ROI size (Figure 6), demonstrating the importance of adequate ROI size for producing reproducible quantitative mineralogy from SEM-EDS mineral mapping even in relatively homogeneous samples.

In addition to the size of the sub-ROI, the shape and alignment of the sub-ROI also impacts on the agreement between SEM-EDS mineralogy and XRD mineralogy as measured by the residuals. The discrepancy between SEM-EDS and XRD mineral abundances is higher and more variable for the laminated sample Qlk6 when the sub-ROIs are divided as a grid than when the sub-ROIs are arranged as columns perpendicular to bedding (Figure 6). The residual values of Qlk6_1 in the gridded sub-ROIs range from 0 to 17.8 wt% and have averages from 1.2 wt% (whole study region) to 2.6 wt% (32 sub ROIs), which is much higher than for the homogeneous sample SX22_1 (0–7.17 wt%, averaging from 0.7 to 0.9 wt%). In contrast, when we use sub-ROIs that are arranged as columns for the laminated sample SX22 (Figure 6C), we find the residuals are similar between Qlk6 and SX22, ranging from 0 to 9.4% with averages from 1.2% (whole study region) to 1.4% (32 sub-ROIs for each sample).

We stress that the greater residuals (i.e., differences between SEM-EDS- and XRD-derived mineralogy) for the gridded ROI pattern versus the column ROI pattern does not reflect incorrect SEM-EDS-based mineralogy results within the gridded ROIs. Instead, the larger residuals reflect grid ROI geometry, size and positioning resulting in the mapped area being compositionally different to the average composition of the bulk sample for vertically heterogeneous samples. Indeed, one of the major advantages of SEM-EDS mineral mapping is the ability to target specific regions of the sample, for example to identify mineralogical differences between adjacent laminae or beds, something which cannot be accessed using bulk mineralogical techniques which only return average sample values. We discuss potential applications of this advantage in Section 4.2.3.

The strong linear correlation between quantitative bulk mineralogy obtained via powder XRD analysis versus SEM-EDS mineral mapping demonstrates the accuracy of SEM-EDS mineral identification and quantification, both for fine-grained calcareous (N = 220, R ² = 0.98) and siliciclastic (N = 48, R ² = 0.98) sedimentary sequences (Figure 3A). This is further supported by the consistent agreement between mineral and grain boundaries observed in BSE images and EDS-based mineral maps (Figures 1, 2, 9). This confirms that the next-generation SEM-EDS techniques implementing the “mixel” approach are able to accurately deconvolve the mixed-phase X-ray volumes that are characteristic of fine-grained lithologies and along grain boundaries, with minimal analyst input or sample-specific modification, representing a major advance over traditional SEM-EDS mineral mapping approaches (Ayling et al., 2012; Johnson et al., 2015; Schulz et al., 2020). However, while the Nanomin^® data reduction algorithm removes much of the manual data processing burden, and therefore limits user introduced error compared to earlier SEM-EDS mineral mapping technologies, this new and improved approach has key points that must be considered, which are highlighted below.

1. Our results show that accurate quantitative SEM–EDS derived “ bulk ” mineralogical data requires careful ROI placement on representative sample sections and consideration of optimal ROI size and geometry. This point highlights the advantages of whole-of sample BSE overview images to inform ROI placement (Section 3.4). Notable outliers (Figure 3A), where data points fall away from the one to one trend line, are confined to four of the twenty ROIs: SX2_1, Qlk7_2, SX9_1 and SX14 (Figure 3A). Three of these ROIs (SX2_1, Qlk7_2 and SX9_1) were specifically placed to characterize areas that are compositionally unusual by comparison to the rest of the sample (i.e., sample regions distinct to the sample average, Figures 1G,H), therefore we would not expected them to fall along the 1:1 trend which is defined by the bulk composition of the samples. By contrast, additional ROIs measured on these three samples specifically positioned to characterize a representative “average” portion of the sample do plot along the 1:1 line (not outliers). Sample SX14 is the sole exception. However, hand sample and SEM inspection of this dolomitic shale shows it to be highly heterogeneous, with compositionally variable laminae as well as irregularly distributed early diagenetic silica cements (silica nodules) and quartz veins. We therefore attribute the outlier data points from this sample to extreme sample heterogeneity, resulting in genuine compositional differences between the SEM-EDS and bulk XRD sample aliquots even after careful ROI placement.

The Nanomin^® SEM-EDS mineral mapping tool presented here facilitates detailed petrographic characterization of fine-grained clastic and carbonate rocks. This represents a fundamental advance, potentially permitting a new approach to a wide range of basic questions in the Earth Sciences that have previously remained difficult to address, as illustrated in the following case studies. We have chosen these examples with the aim of illustrating a broad range of novel applications, and hope to encourage broader uptake of sedimentary petrography as a versatile, complementary tool in the Earth scientist’s arsenal.