Classification of boiling regimes, fluids, and heating surfaces through deep learning algorithms and image analysis

Precise monitoring and forecasting of boiling dynamics are crucial for ensuring reliability near critical conditions in thermal systems. This study applies Convolutional Neural Network (CNN) algorithms for image classification to identify boiling phenomena under different operating conditions, including single-phase flow, nucleate boiling, and pre-critical heat flux states. The dataset used to train the models was obtained from a Flow-Boiling experimental setup with Joule effect heating, a configuration less explored in this regard compared to pool boiling. Additionally, the study also classifies the working fluid (water, ethylene glycol-water mixture, and hydrofluoroether) and the heating plate (non-textured or micro-textured surfaces). Four CNN architectures (AlexNet, ResNet, InceptionNet, and a standard CNN) were evaluated using confusion matrices and performance metrics including precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). The performance was also compared to a previous methodology involving downsampling, Principal Component Analysis, and a Support Vector Machine. This previously reported method achieved MCC values of around 58 % for boiling regime classification and 61 % for fluid classification. In contrast, advanced CNN models demonstrated significantly superior performance. AlexNet achieved MCC values of 97 % and 96 % for boiling regime classification and excelled in fluid type classification, with MCC values of 98 % and 99 % for training and testing, respectively. For plate type classification, InceptionNet achieved an F1-score of approximately 98 %. These findings highlight the effectiveness of CNN algorithms in accurately classifying boiling phenomena, offering robust tools for analyzing and monitoring flow boiling systems through direct visualization.