Interpretable machine learning for quantitative parameter importance and operational thresholds in PWR accident prediction

The "black box" nature of machine learning models hinders trust and transparency in nuclear safety systems, where interpretability is critical. This study introduces an explainable CatBoost-SHAP framework for accident prediction in pressurized water reactors (PWRs). Leveraging CPR1000 reactor simulator and Optuna-optimized CatBoost, the model achieved high accuracy (R² > 0.999, MAPE <1 %) on small break loss-of-coolant accident (SBLOCA) datasets for both hot-leg and cold-leg scenario, outperforming XGBoost and LightGBM. SHAP analysis identified key thermal-hydraulic drivers (e.g., steam generator (SG) pressure <6.74 MPa, wide-range downcomer level <−1.4 %) and uncovered nonlinear interactions among multi-loop variables, consistent with reactor physics. The framework's dual capability - high predictive precision and mechanistic interpretability - enables operators to validate decision pathways and prioritize safety thresholds. By bridging the gap between opaque AI and nuclear safety demands, this work provides practical guidelines for real-time diagnostics and proactive accident mitigation in PWRs.