Machine-learning-based heat transfer and pressure drop model for internal flow condensation of binary mixtures

Abstract

Machine learning regression models were developed to predict the heat transfer coefficient and frictional pressure gradient during condensation of three zeotropic mixtures (ethane/propane, R245fa/pentane, and R410A) in micro- and macro-channels (0.76 mm < D < 14.45 mm). 1032 data points were used to train, test, and validate four machine learning models: support vector regression (SVR), random forest regression (RFR), gradient boost (GB), and artificial neural networks (ANN). For data not used for training, SVR predicted the apparent Nusselt number the best, with a mean absolute percent error (MAPE) of 5.0%, while GB predicted the two-phase friction factor the best with a MAPE of 5.5%. To determine the impact of the additional heat and mass transfer resistances in zeotropic condensation, Shapley analysis was performed on the models to weigh the importance of all dimensionless parameters used to predict the heat transfer and pressure drop. Overall, the three most important parameters to predict heat transfer were the Reynolds number of both phases and a dimensionless temperature glide; while the most important parameters to predict pressure drop were the Bond number, Weber number, and vapor-phase Reynolds number. The importance of mass transfer is discussed along with the applicability of other simple mixture models.