Physics-informed machine learning framework for boiling heat transfer prediction of dielectric fluids

Existing boiling correlations provide theoretical footings, but their applicability remains limited by specific parameter spaces due to inherent nonlinear interactions between two-phase flow, mass and heat transfer behaviors. While data-driven machine learning shows promising prediction accuracy, its extrapolation capability heavily relies on the dataset quantity and lacks the mechanistic interpretability. To overcome these issues, this study proposes a physics-informed machine learning model by combining a physics-based correlation with machine learning approach. The hybrid framework achieves superior prediction accuracy. Specifically, the theoretical correlation lays the groundwork for domain knowledge, while the machine learning captures the explicit information from knowledge-predicted targets. A detailed study is performed using consolidated dataset (907 datapoints from 24 literature resources) with respect to dielectric fluids. Based on this, a new boiling heat transfer correlation for dielectric fluids is proposed, which outperforms the original correlations and provides improved prior knowledge. The fully data-driven models are also comprehensively evaluated, showing remarkable data quality dependence. Results suggest that the modified correlation combined with CatBoost regressor realizes the desired predictive performance in comparison to standalone models. Additionally, the physics-informed machine learning model exhibits robust generalization across different dataset quantities because modified correlation can offer baseline knowledge to reduce the prediction variance.