A Co-designed deep learning framework embedding physical insights for precise prediction of critical heat flux in microchannel flow boiling

• A comprehensive CHF database for structure-enhanced microchannels flow boiling heat transfer was established. • A deep learning framework (DNN & d -KAN) accurately predicts CHF with MARE ≤ 7.4 %. • Adaptive weighted loss improves CHF prediction under low-power operation conditions. • Simplifying input to basic parameters maintains high predictive accuracy and robustness. • SHAP interpretability validates the physical rationality of AI-based CHF prediction. The continued miniaturization of high-power electronic devices has exacerbated the associated thermal management challenges. Among various cooling technologies, microchannel flow boiling heat transfer has gained significant attention due to its high thermal dissipation efficiency and ease of integration. Critical heat flux (CHF), as the upper limit of boiling heat transfer performance, is therefore a key parameter for the reliable design of microchannel heat exchangers. However, the complex physical mechanisms and strong coupling among operating conditions, surface structures, and thermophysical properties significantly limit the predictive accuracy of traditional CHF correlations, particularly for structure-enhanced microchannel flow boiling. In this study, a comprehensive experimental database covering a wide range of operating conditions and working fluids is constructed, and a deep learning–based prediction framework is developed by incorporating physically motivated input features related to CHF triggering mechanisms. The final optimized deep neural network (DNN) model achieves high prediction accuracy, with a mean absolute relative error (MARE) below 7% and a coefficient of determination (R²) exceeding 0.99, demonstrating a substantial improvement over conventional empirical correlations. Further simplification of the input features to basic geometric, operating, and thermophysical parameters preserves robust predictive performance, facilitating practical engineering application. In addition, an adaptive weighted loss function is introduced to enhance prediction accuracy in the low-CHF regime, addressing a long-standing challenge in multi-fluid CHF prediction. SHAP-based interpretability analysis confirms that the learned feature importance is consistent with established physical understanding of flow boiling CHF. Overall, the proposed framework provides an accurate and physically consistent tool for CHF prediction in structure-enhanced microchannels, offering practical guidance for microchannel thermal design and advanced cooling applications.