Universal Physical Descriptors of Boiling Heat Transfer from Unsupervised Learning

Understanding the mechanisms during boiling has been challenging due to the high dimensionality and stochasticity of the bubble dynamics. Conventional analysis of boiling images and data are limited to existing knowledge. Unsupervised machine learning is a powerful tool to uncover physical insights into the bubble dynamics during boiling without human supervision. This study demonstrates the universality of physical descriptors of boiling heat transfer extracted from pool boiling experimental images using principal component analysis (PCA), an unsupervised dimensionality reduction algorithm. The physical descriptors are interpretable when compared with conventional parameters such as the size and count of bubbles as well as vapor area fraction for a wide range of heater surface-working fluid pairs. The time-series principal components (PCs) are analyzed, where the correlation between dominant frequencies of the PCs and the heat flux is discovered, linking heat transfer performance with bubble nucleation and coalescence events. The approach works well for both low and high surface tension fluids, as well as diverse engineered heat transfer surfaces.