Generalizable physical descriptors of pool boiling heat transfer from unsupervised learning of images

Abstract

Boiling processes are notoriously difficult to analyze via visual inspection due to the complex interactions between the vapor bubbles and the surface. Unsupervised machine learning (ML) is a powerful tool to uncover physical insights into the bubble dynamics during boiling from image data. In this study, principal component analysis (PCA), an unsupervised dimensionality reduction algorithm, is used to extract new physical descriptors of boiling heat transfer from pool boiling experimental images without any labeling and training. Experiments are conducted with different working fluids and heater surfaces to investigate the effect on the bubble morphology and subsequently on the physical descriptors identified through unsupervised ML. The dominant frequency and amplitude deduced from the Fourier transform of the time series of the first principal component (PC) are compared against physical parameters such as bubble size, bubble count, and vapor area fraction. The new physical descriptors derived from PCA show a positive correlation with conventional parameters related to bubble morphology, as demonstrated by linear regression analysis. Pearson Correlation Coefficients further confirm the strong correlations between dominant amplitude and both bubble size and vapor area fraction, as well as between dominant frequency and bubble count. These strong correlations hold across multiple different working fluids (water and HFE 7100) with different heater surfaces (plain and microstructured surfaces made of copper and silicon materials), demonstrating the potential for these extracted physical descriptors to generalize and act as a surrogate to conventional physical descriptors. This unsupervised learning approach offers a robust alternative to traditional pool boiling analyses or supervised ML approaches that rely on time-consuming manual labeling involving bubble identification and segmentation.