Machine learning for identifying characteristics of isolated, clustered, and pulsed vapor bubbles on a heated surface under non-stationary boiling conditions

This paper presents an automated system for analyzing high-speed video of non-stationary nucleate boiling on an opaque steel surface. The method leverages the DenoiSeg deep learning network for robust bubble segmentation under challenging conditions (reflected light, optical distortions) and introduces an algorithm for tracking bubbles and calculating time-dependent characteristics. The system identifies and classifies bubbles into three types (isolated, clustered, and pulsating) to extract the essential boiling parameters, including nucleation site density, surface area fraction, maximum diameter, and nucleation frequency. Validation against manual key frame analysis confirms the system's accuracy. The results not only verify the significant prevalence of clustered and pulsating bubbles but also, thanks to extensive data processing, reveal trends hidden by stochastic noise, such as the growth of the maximum diameter of clusters with increasing surface temperature. The developed tool provides a reliable foundation for building predictive heat transfer models for non-stationary boiling regimes