Bayesian-optimized, feature-augmented deep ensemble for physics-guided critical heat-flux prediction with uncertainty quantification

Accurate prediction of the critical heat flux (CHF) is a crucial design and safety consideration for a wide range of high-performance thermal systems, including water-cooled nuclear reactors. Traditional predictive tools, such as empirical correlations and look-up tables, often lack accuracy when extrapolated or at different interpolation regions. To overcome these limitations, this work introduces a novel physics-guided, optimized deep-ensemble framework for robust CHF prediction with comprehensive uncertainty quantification. Our approach first expands the model’s inputs by augmenting base thermal-hydraulic parameters with physics-based features derived from established correlations. This feature engineering injects domain knowledge, constraining the solution space and promoting convergence to physically plausible solutions. Furthermore, we employ a sophisticated hyperparameter optimization strategy, combining a Sobol sequence with Bayesian optimization, to systematically select a diverse and high-performing set of neural networks for the ensemble. The resulting physics-guided ensemble demonstrates superior performance across all metrics compared to a baseline ensemble, a standard look-up table, and a benchmark neural network. The model produces smoother, more physically consistent predictive trends and provides reliable uncertainty estimates. This framework offers a powerful and broadly applicable tool for CHF prediction, enabling higher-fidelity safety margins and the design of more efficient and reliable thermal management systems. • Developed a novel physics-guided deep-ensemble framework for CHF prediction. • Integrated Sobol sequences and Bayesian optimization for rigorous model tuning. • Achieved 5%–7% error reduction using correlation-based feature augmentation. • Outperformed the 2006 Groeneveld look-up table and OECD benchmark models. • Established robust aleatoric and epistemic uncertainty quantification.