Harnessing multi-source data to develop universal machine learning models for saturated flow boiling heat transfer in mini/macro channels

Reliable knowledge of Heat Transfer Coefficient (HTC) under flow boiling conditions underpins the optimization of thermal systems where compact, high-performance heat exchangers are required. This study develops and evaluates Machine Learning (ML) models for predicting HTC in saturated flow boiling within mini and macro channels. A large and diverse experimental database containing 24,614 data points from 120 sources is compiled to train and validate four ML regression approaches for construction of universal predictive models. The results show that all models outperform existing correlations, with mean prediction errors of around 4 % for the majority of models. Tests on excluded data subsets demonstrate reasonable generalization capability, although performance declines when operating conditions differ significantly from the training data. Analysis of these cases highlights the importance of data representativeness and provides insights for guiding future experimental efforts. Overall, the proposed ML-based framework offers a more accurate and versatile tool for HTC prediction compared to conventional methods. • Developed universal ML models to predict heat transfer coefficient in boiling flows. • Compiled a large database with 24,614 points across 120 sources and diverse conditions. • Compared ANN, GBR, XGBR and RFR algorithms, achieving MARD below 4 % for top models. • Models show reasonable extrapolation with ∼30 % error on unseen data subsets. • Identified data and target discrepancies as key factors limiting model generalization.