Work

Large-scale vision-based interfacial characterization for heat transfer prediction in horizontal mini-channel flow boiling

Year: 2026

Type: article

Abstract: Flow boiling is a promising solution for high heat flux thermal management, but existing models struggle to capture the strong coupling between local phase distribution, flow regime, and heat transfer coefficient (HTC). In this work, we combinehigh-speed visualization, advanced computer vision, and machine learning to build a regime-aware framework for horizontal mini-channel flow boiling under terrestrial conditions. A large dataset of 107 cases over broad ranges of inlet mass flux, wall heat flux, and inlet subcooling is acquired in a single mini-channel, and more than 3.8×105 images are processed with benchmarked instance-segmentation models, from which a refined YOLOv8-based pipeline is selected. The vision model yields spatially resolved,time-averaged interfacial features, including streamwise void-fraction and dry-fraction distributions, which reveal systematic trends with operating conditions and multiple flow regimes. These image-derived void-fraction profiles are then used to evaluate severalstandard void-fraction correlations, highlighting regime- and condition-dependent strengths and limitations. Finally, hybrid neural networks combining scalar operating parameters with one-dimensional axially varying interfacial distributions are developed:a quality-informed baseline using streamwise quality, and a region-resolved model using vision-based void- and dry-fraction distributions. The region-resolved model predicts local HTC with mean absolute percentage errors of about 14–15%, while offeringenhanced physical interpretability, demonstrating that vision-based interfacial distributions can effectively replace explicit quality profiles in HTC prediction and support more reliable, data-driven flow-boiling design.

Source: Applied Thermal Engineering

DOI: 10.1016/j.applthermaleng.2026.132079