Physics-constrained machine learning for scenario-based prediction of far-distance radionuclide attenuation at the Lufeng Nuclear Power Plant

Radiological consequence assessment beyond directly simulated receptor distances remains difficult because rapid emergency tools are designed primarily for timely screening rather than long-range prediction. This study developed a physics-constrained neural surrogate (PCNS) to reproduce and cautiously extend RASCAL-derived radiological consequence patterns for three hypothetical severe accident scenarios at the AP1400-based Lufeng Nuclear Power Plant, China. The scenarios comprised progressive core melt with containment leakage, steam-generator tube rupture, and total containment bypass with reactor-vessel failure, each evaluated at 60 and 100 m release heights using site-specific meteorological inputs. The modelling dataset included source terms, release rates, dose endpoints, deposition, time-integrated air concentration, and radionuclide-specific outputs for I-131, Cs-134, Cs-137, Sr-90, and Te-132. Random Forest, Gradient Boosting, XGBoost, and a data-driven neural network were used as benchmarks. The proposed surrogate incorporated monotonicity and smoothness penalties to reduce non-physical increases and unstable oscillations with distance. XGBoost achieved the strongest performance among the conventional models, while the smoothness-constrained surrogate showed the closest agreement with withheld RASCAL-derived reference values, attaining a distance-aware test R 2 of 0.9991 and an RMSE of 0.0913 in log 10 space. Ensemble spread increased with extrapolation distance, indicating declining confidence outside the directly simulated domain. The framework is best interpreted as a rapid, uncertainty-aware screening tool for scenario comparison and protective-action planning, not as an independently validated atmospheric dispersion model.