2026-09-24T13:31:06Zhttps://api.repository.cam.ac.uk/server/oai/request

oai:www.repository.cam.ac.uk:1810/2217252025-12-19T19:13:29Zcom_1810_34586com_1810_256064col_1810_205358

Mechanical properties of materials for fusion power plants 10.17863/CAM.14233 Forsik, Stéphane Alexis Jacques Fusion power is the production of electricity from a hot plasma of deuterium and tritium, reacting to produce particles and 14 MeV neutrons, which are collected by a cooling system. Their kinetic energy is transformed into heat and electricity via steam turbines. The constant ux of neutrons on the rst wall of the reactor produces atomic displacement damage through collisions with nuclei, and gas bubbles as a result of transmutation reactions. This leads eventually to hardening and embrittlement. Designing a material able to withstand such intensity of damage is one of the main aim of research in the eld of controlled fusion. In the past decades, many experiments have been carried out to under- stand the formation of radiation{induced damage and quantify the changes in mechanical properties of irradiated steels, but the lack of facilities prevents us from testing candidate materials in a fusion{like environment. Modelling techniques are utilised here to extract information and principles which can help estimate changes in steels due to damage. The elongation and yield strength of various low{activation ferritic/ martensitic steels were modelled by neural networks and Gaussian processes. These models were used to make predictions which were compared to exper- imental values. Combined with other techniques and thermodynamic tools, it was possible to understand the evolution of the mechanical properties of irradiated steel, with a particular focus on the role of chromium and the roles of irradiation temperature and irradiation dose. They were also used to extrapolate data related to ssion and attempt to make predictions in fusion conditions. A set of general recommendations concerning the database used to train the neural networks were made and the usage of such a modelling technique in materials science is discussed. An attempt to optimise the performance of neural networks by suppress- ing some random aspects of the training is presented. Models of the elon- gation, yield strength and ductile-to-brittle transition temperature trained following this procedure were created and compared to classical models. University of Cambridge 2009-08 Thesis Doctoral Doctor of Philosophy (PhD) en http://www.dspace.cam.ac.uk/handle/1810/221725 https://www.repository.cam.ac.uk/handle/1810/221725 https://www.repository.cam.ac.uk/bitstreams/f0d0137b-cd3b-4e6f-87ae-84915492e102/download 8a1a2c22ffaec9b1024fb31998c7b4f7 https://www.repository.cam.ac.uk/bitstreams/77e4d76c-077d-40a6-b845-661e42f8c85f/download 0df08cf47b329d3251994bcdfa42b5a6 https://www.rioxx.net/licenses/all-rights-reserved/ 2009