Forward ensemble machine learning for grouped source term prediction in nuclear emergency response

Major nuclear accidents reveal critical limitations in rapid source term prediction during emergency response. Conventional high-fidelity simulation approaches are computationally expensive and unsuitable for near-real-time decision-making. This study develops a machine learning surrogate framework to predict grouped radionuclide source terms-noble gases, iodine, and particulates-under a hypothetical containment-bypass scenario for the CAP1400 reactor. Using 3000 RASCAL-generated simulation samples for training and testing, three ensemble models were evaluated: LightGBM, Stacking, and CatBoost. LightGBM achieved the best overall performance on unseen test data (R 2 = 0.714) with near-instantaneous inference (∼ 0.001 s). Noble gases showed the highest predictive accuracy (test R 2 = 0.762), whereas particulates remained the most challenging target (R 2 = 0.58–0.74). SHapley Additive exPlanations (SHAP) analysis identified mass number, core uncovered time and particulate concentration as dominant predictive features. The results demonstrate the potential of machine learning surrogate models to provide rapid and reliable source term estimates for near-real-time nuclear emergency response.