Machine learning assessment of the importance of unirradiated yield strength as a variable in embrittlement trend forecasting

Abstract

This paper presents an investigation into the possible influence of pre-irradiation hardening of RPV steel on the transition temperature shift, ΔT_41J. Using the ASTM PLOTTER-22 database supplemented with unirradiated yield stress, YS(u), data the study uses machine learning regression algorithms to construct a predictive model that accounts for YS(u) alongside more well-established predictor variables (e.g., copper, fluence, …). The Gradient Boosting algorithm emerged as the most efficient, with performance metrics R² = 0.89 ± 0.02 and root-mean square error (RMSE) = 11.2 ± 0.7 °C. Comparative analyses via bootstrapping underscore the beneficial effect of incorporating YS(u) as a regressor, resulting in a RMSE reduction by 7 % and R² improvement of 15 %. Feature interpretation techniques demonstrate that the significance of YS(u) is comparable to elements like nickel and irradiation temperature and above others such as manganese, phosphorus, or the product form of the steel. The revealed trend — higher YS(u) corresponding to lower ΔT_41J — and the lack of significant interactions between YS(u) and the chemical composition, supports the roughly independent role of YS(u). These results underscore the value of incorporating YS(u) as a predictor variable for irradiation embrittlement.