Unit-aware physical embedding transformer for critical heat flux prediction and extrapolation at boiling crisis boundaries

Critical heat flux (CHF) is a key thermal limit for boiling heat transfer and reactor safety, but data-driven CHF models often suffer from non-physical extrapolation near boiling crisis boundaries. In this study, a unit-aware physical embedding Transformer is proposed for CHF prediction. The model integrates learnable unit embeddings, unit-interaction features, multi-scale extraction and physics-informed regularization to preserve the physical semantics and scale consistency of thermal–hydraulic variables. The model was evaluated on 24,579 samples from the public CHF tube database under full-range testing and two extrapolation scenarios: high-pressure holdout and high-quality holdout. In the full-range test, the proposed model achieved a root mean squared error (RMSE) of 213.91 kW m−2 and a coefficient of determination (R2) of 0.9818. Under extrapolation conditions, the proposed framework maintained stable performance in high-pressure regimes and achieved the best overall accuracy in the high-outlet-quality dryout-dominated regime, with an RMSE of 327.87 kW m−2 and an R2 of 0.7850. Unit-perturbation tests further showed that the model reduced systematic prediction drift by more than 60%. Interpretability analyses indicated physically consistent responses associated with mass flux, pressure and outlet quality. These results demonstrate that embedding unit semantics and physical constraints into Transformer-based CHF modeling can improve the robustness and credibility of thermal-limit prediction under extrapolative boiling crisis conditions.