Multi-head attention network for prediction of irradiation embrittlement in reactor pressure vessel steels

The accurate prediction of irradiation embrittlement in reactor pressure vessel (RPV) steel remains challenging. Here, we develop a multi-head attention network (MHANet) and target the service-relevant shift in ductile-brittle transition temperature (ΔDBTT) co-varying with temperature, flux, and fluence. MHANet treats each input column as a token, learns feature dependencies via attention, and yields accurate predictions with transparent attribution. Trained/validated on 1266 IVAR+NUREG samples, MHANet achieves MAE of 5.63 ∘C(IVAR) and 16.21 ∘C (NUREG), and outperforms a baseline multilayer perceptron model. The learned dependencies are found to be physically consistent with the coupled effects by which ΔDBTT is governed, including point defects, solute-defect interactions (particularly involving Cu, Ni, and P), and irradiation-induced precipitation. (i) a higher temperature mitigates embrittlement; (ii) at a fixed fluence, the reduction in ΔDBTT with increasing flux is attributed to the competition between steady-state defect concentrations and the available evolution time; (iii) threshold-like synergies emerge-ΔDBTT increases sharply when Ni ≈ 0.9–1.1 wt.% and Cu (≥0.10 wt.%) or P is elevated. These results provide a coherent path from data-driven prediction to mechanism-aware interpretation and actionable criteria for RPV life assessment and alloy design.