This study develops a stacking ensemble machine learning model to predict ductile-to-brittle transition temperature (DBTT) in irradiated Reduced-Activation Ferritic-Martensitic (RAFM) steels. Using a dataset of 490 irradiation cases, the model integrates XGBoost, Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Multi-Layer Perceptron (MLP), achieving a high predictive accuracy (R² = 0.96) and outperforming individual models. The results highlight the significant influence of irradiation dose and temperature on DBTT. Beyond 30 dpa, defect accumulation causes a sharp DBTT increase, while irradiation temperature exhibits a nonlinear effect, peaking at 150-300°C due to radiation-enhanced precipitation and declining above 350°C as defect recovery improves ductility. Additionally, alloying elements play a crucial role: Ta and W help mitigate embrittlement, moderate Cr (4–6 wt.%) increases DBTT, and higher Cr levels (>6 wt.%) reduce it at elevated temperatures. SHAP analysis reveals that W is particularly effective in reducing embrittlement in alloys with moderate Cr (6–9 wt.%) and low Ta, while higher Cr concentrations (>6 wt.%) help stabilise DBTT at elevated temperatures. To enable practical alloy design, a genetic algorithm was combined with the model to optimise steel compositions under defined irradiation conditions (200–350°C, 10 dpa). The approach successfully identified candidate alloys with predicted DBTT values below 50°C. This study provides a robust predictive framework for understanding and optimising DBTT in irradiated RAFM steels, offering valuable insights into their performance in next-generation nuclear reactors.

Published in Journal of Nuclear Materials, Vol. 615 (2025), Article 155984.