12th International Symposium on Symbiotic Nuclear Power ( ISSNP-2026)

Predicting critical heat flux (CHF) with both high accuracy and high interpretability remains a central challenge in nuclear thermal-hydraulics, especially as artificial intelligence (AI) increasingly complements experiments and mechanistic models. While modern deep-learning architectures have demonstrated strong predictive capability, their black-box nature limits trust and adoption in safety-critical reactor applications. Conversely, symbolic regression offers transparent functional forms but often sacrifices accuracy when dealing with high-dimensional, nonlinear CHF behavior. This study aims to bridge this gap by developing a hybrid framework that integrates advanced deep-learning models with interpretable symbolic-regression correlations to achieve both high predictive performance and physical transparency. We first revisit CHF-related input-parameter engineering by systematically evaluating hydrodynamic, thermal, and geometric features drawn from major databases covering a wide range of pressures, flow regimes, and heating conditions. Transformer-based networks and gradient-boosting ensemble models are trained to establish an accuracy benchmark using optimized feature combinations. Subsequently, PySR-based symbolic regression is employed to derive explicit analytical correlations. A hybrid workflow is introduced in which deep-learning models provide feature-importance guidance, enabling symbolic regression to identify simplified yet physically meaningful expressions while maintaining high accuracy. Results show that the Transformer model yields state-of-the-art performance with markedly reduced prediction error compared with existing correlations. More importantly, symbolic regression—when constrained by deep-learning-guided feature selection—produces compact equations that outperform classical CHF correlations and closely match the accuracy of deep networks. The derived expressions explicitly capture dominant parameter sensitivities, such as mass flux, pressure, inlet subcooling, and heated-length effects, consistent with known boiling physics. This synergy between accuracy and interpretability demonstrates that CHF can indeed be predicted reliably without relying solely on black-box models. The proposed hybrid approach highlights a promising direction for trusted AI in nuclear thermal-hydraulics and reactor safety, offering both practical predictive capability and transparency necessary for licensing, analysis, and design applications.
 

Critical Heat Flux (CHF); Deep Learning; Symbolic Regression; Interpretable AI; Boiling Heat Transfer; Nuclear Thermal-Hydraulics