Image driven deep learning based compact model to predict critical heat flux in direct immersion cooling via pool boiling

This paper introduces a concise model, based on convolutional neural networks (CNNs), to predict the performance of two-phase boiling heat transfer systems by leveraging data acquired through high-speed visualization techniques. The proposed approach utilizes high-speed digital imaging to meticulously capture and analyze bubble behavior on a heating component. Through this, a quantitative relationship is established between physical data and visual patterns, offering detailed insights into dynamic parameters such as bubble growth, departure, coalescence, and the static parameter of bubble diameter. The Convolutional Neural Network (CNN) is trained using visual patterns and physical data obtained from the visualization technique. Two distinct frameworks are employed for training: the first framework (A1) deploys Mask-R-CNN to extract physical data from images and utilizes a machine learning model trained on experimental data for interpolation. The second framework (A2) focuses on characteristic categorization, classifying test images into specific heat flux categories based solely on visual patterns derived from convolution operations. The study aims to quantify the relative weights assigned to the categorization and interpolation algorithms. Furthermore, this research endeavors to tackle a critical challenge encountered by AI models which is quantification of similarity in the training data to ensure optimal prediction accuracy.