Physics-Driven Learning for Two-Phase Heat Transfer

초록

Two-phase heat transfer, including condensation and boiling, is critical in applications such as energy systems, electronics cooling, and thermal management. However, accurately modeling these processes remains challenging due to their complex, nonlinear interactions. This study employs physicsinformed machine learning (PIML) to enhance predictive accuracy and generalizability. For condensation, a novel degradation Nusselt model integrated with a PIML framework significantly improved extrapolation accuracy, reducing the mean absolute percentage error (MAPE) from 21.63% (purely datadriven models) to 11.22%. For boiling, critical heat flux (CHF) predictions were refined by incorporating mechanistic insights and tree-based models, reducing MAPE from 20.41% to 12.7% in extrapolation scenarios. This work bridges data-driven modeling with physics-based constraints, ensuring predictions align with fundamental heat transfer principles. The findings establish a robust framework for next-generation thermal management systems, with opportunities for future research to expand datasets and integrate additional physical constraints for broader applicability.

키워드