Advanced machine learning analysis of radiation hardening in reduced-activation ferritic/martensitic steels

Abstract

This study uses advanced machine learning models to investigate the radiation hardening behaviour of reduced activation ferritic martensitic (RAFM) steels. An extensive dataset spanning nearly four decades (1985 to 2024) and covering various steel series, including Eurofer97, F82H, T91, OPTIFER, JLM, JLF, and CLAM, was extensively analysed. Multiple models, including Gradient Boosting Decision Trees (GBDT), XGBoost, Random Forests (RF), ResMLP, and One-Dimensional Convolutional Neural Networks (1D-CNN), were employed with hyperparameter optimisation to maximise predictive accuracy. Among these models, GBDT achieved the highest accuracy (R²: 0.87). The findings reveal significant impacts from elements like Ta, W, and Cr, as well as test temperature and irradiation dose. Radiation hardening peaks at 315 °C due to increased dislocation loops and precipitates but declines above 375 °C as these features diminish and martensitic laths recover, softening the steel. The hardening response to radiation dose shows an increase up to 20 dpa, a slight decrease between 20–35 dpa, and stabilising thereafter. Additionally, W and Cr enhance radiation hardening up to 375 °C, with Cr exhibiting a stronger effect, while Ta is observed to mitigate hardening. These insights contribute to a deeper understanding of radiation effects on RAFM steels, offering a predictive framework for material design and optimisation in nuclear environments. This work highlights machine learning as a powerful tool for advancing materials science and enhancing predictive capability for radiation behaviour in steels.