Performance evaluation and adaptive variable-order method of thermal-hydraulic simulations models based on random forests

High-order separated flow models used in nuclear thermal–hydraulic simulations often fail under complex transients. To enhance stability and engineering applicability, this study proposes a machine learning-based method for model performance evaluation and adaptive order adjustment. A random forest classifier, optimized by probability calibration and threshold adjustment, was trained using a dataset of M310 and HPR1000 reactors. Borderline- Synthetic Minority Oversampling Technique was used to address class imbalance and improve detection of unstable cases. The method achieved 95.64 % recall, 92.31 % precision, and 0.005 s average evaluation time. The final model, integrated as a C++ dynamic-link library, enables early warning of calculation failures within simulation environments. This approach significantly improves the robustness and intelligence of thermal–hydraulic simulations, filling a gap in automated failure prediction.