Towards data-driven boiling models using a mixture-of-experts approach

Accurate prediction of boiling heat transfer remains challenging because nucleate boiling involves complex, multi-scale interactions and existing heat flux partitioning models often rely on empirical or semi-empirical sub-models with limited validity. This paper proposes a data-driven mixture-of-experts framework that combines multiple semi-mechanistic and semi-empirical boiling models through a neural gating network trained on subcooled flow boiling data. Each expert represents a distinct combination of heat flux partitioning formulation and bubble dynamics sub-models, while the gating function adaptively weights their contributions using non-dimensional flow and thermophysical features. The approach improves global predictive accuracy compared with individual boiling models while retaining interpretability through explicit model weights.

Proceedings of the 18th International Heat Transfer Conference