Two-stage machine learning framework for investigating vapor bubble dynamics in nucleate flow boiling

Nucleate boiling is a highly efficient heat-transfer mechanism involving complex bubble dynamics that still remains difficult to predict. Traditional high-speed videography based experimental approaches are time and resource-consuming and generate large image datasets that demand automated processing. This study proposes a novel two-stage machine-learning framework for vapor bubble dynamics in nucleate flow boiling: Stage I employed a comparative evaluation of seven deep learning segmentation models (including DeepLabV3+, U-Net, and others) using three different loss functions, namely binary cross-entropy (BCE), Dice loss, and combined BCE-Dice loss, to precisely segment vapor bubble boundaries; Stage II involved a comparative analysis of four machine learning algorithm families and their variants to predict critical bubble dynamic parameters, including aspect ratio (AR), upstream and downstream contact angles ( θ u , θ d ), equivalent diameter ( d e q ), and base diameter ( d b ) from experiments in a vertical channel under varying subcooling (2–20 K), flow rate (6–40 lph), heat input (24–47 W), and normalized cycle time (0–1). Results demonstrated that DeepLabV3+ with combined BCE-Dice loss achieved superior performance (validation IoU 0.982, training IoU 0.977) with minimal parameter errors (8.28% for θ u , 9.71% for θ d , 1.49% for d e q , 2.27% for AR, and 4.47% for d b ) in Stage I. Moreover, the combined BCE-Dice loss function consistently outperformed individual loss functions across most segmentation models. The Bagging Regressor was the optimal predictor in Stage II (mean R 2 of 0.858; average error of 2.4, 5.4, 5.7, 6.7, and 8.2% for AR, θ u , θ d , d e q , and d b , respectively). This integrated approach automates both segmentation and bubble-parameter prediction, dramatically reducing manual processing and enabling rapid, data-driven design of boiling-based thermal systems.