Advancing nuclear energy safety via hybrid stacking models: Integrating physics-informed machine learning and traditional AI for critical heat flux prediction

Critical heat flux (CHF) sets the primary thermal-safety limit in water-cooled reactors, yet conventional look-up tables (LUT) and mechanistic correlations degrade when extrapolated to high-pressure, high-mass-flux conditions. This study embeds a Groeneveld-type LUT in a physics-informed machine-learning (PIML) framework and applies stacked generalization to correct its residuals. Four Bayesian-optimized base learners generate hybrid predictions whose errors feed single (St1) and double-layer (St2) stacks. Training on 24,579 OECD-NEA/NRC measurements spanning 0.1–20 MPa, 8–7964 kg⸱m −2 ⸱s −1 and 2–16 mm channels, the best five-input St1 model attains MAE = 0.095, RMSE = 0.167, rRMSE = 9.24 % and R 2 = 0.989 under five-fold cross-validation. Transfer-learning tests on unseen operating maps confirm strong extrapolation, while pruning a weak branch further enhances robustness. The resulting hybrid-stacking tool is fast, interpretable, and highly accurate, offering enlarged thermal margins for advanced reactor design and real-time safety monitoring.