Machine learning boiling prediction: From autonomous vision of flow visualization data to performance parameter theoretical modeling

Flow boiling is a highly efficient configuration for meeting the high heat dissipation demands of thermal management systems. However, the complex physics of two-phase flow has hindered its broader application, especially in terms of quantifying visual information. Recent advancements in machine learning vision tools have revolutionized the analysis of phase change phenomena by enabling the digitalization of physically meaningful features such as void fraction, vapor-liquid interfacial behaviors, and liquid-solid wall wetting front areas en masse. In this study, we systematically investigate two-phase models that compute void fractions, heat transfer coefficients, and critical heat flux using live bubble data streams under microgravity. The collected empirical bubble data is used to supplement and validate traditional control-volume-based theoretical modeling approaches. Void fraction data is first validated with analytical frameworks. This is followed by void fractions and wetting front areas being used to improve correlations predicting heat transfer coefficients. This work showcases the potential of using a new machine learning-based strategy to accelerate scientific formula discovery through the extraction of multi-level and physically meaningful features.