In this study, we develop a correlation-informed neural networks (CINNs) framework for predicting flow boiling heat transfer coefficients (HTC) in mini/micro-channels, addressing key limitations of purely data-driven artificial neural networks (ANNs) models and traditional empirical correlations. The proposed framework integrates consolidated experimental data with an established universal correlation through a physics-informed loss function, enabling the model to leverage both data-driven flexibility learning and physically grounded constraints. A comprehensive database of around 17,000 data points, along with four representative universal correlations, is used to train and evaluate the CINNs model. A physics-weighting parameter is introduced to modulate the balance between measured data and embedded physical knowledge. Model evaluations demonstrate that the CINNs framework achieves markedly improved prediction accuracy relative to standalone empirical correlations and exhibits enhanced robustness compared to the purely data-driven ANNs model, particularly when the training dataset is limited. Uncertainty quantification analyses further reveal the sensitivity of the CINNs performance to the underlying physics model and the amount of available training data. In addition, tests on unseen fluids show that the CINNs model provides substantially better extrapolation capability than the ANNs model, highlighting the benefits of incorporating physics-based structure into the learning process. Finally, the CINNs framework is leveraged to refine an existing correlation (Kim and Mudawar) by learning consistent parameter adjustments, yielding an ≈ 4% improvement in predictive accuracy. Taken together, these results demonstrate that the CINNs framework offers a powerful and data-efficient approach for modeling flow boiling heat transfer in compact channels, providing a significant advancement over conventional methods.
Phan et al. (Fri,) studied this question.