A universal correlation for flow condensation heat transfer in horizontal tubes based on machine learning

Abstract

Accurate and universal prediction of in-tube condensation heat transfer coefficients (HTCs) is vital for designing compact condensers. This study presents machine learning (ML) methods for predicting flow condensing HTCs inside horizontal tubes based on an assembled database. The database contains 6064 experimental data of 28 pure fluids and covers broad operating conditions. Furthermore, the database is serviced to train and evaluate five ML models based on K-nearest neighbors (KNN), artificial neural network (ANN), convolutional neural network (CNN), random forest (RF), and Extreme gradient boosting (XGBoost) algorithms. The designed ML models show excellent predictive performances for 1213 test data points with the best mean absolute relative deviation (MARD) of 5.82% achieved by CNN and the coefficient of determination (R2) of 0.98 or higher for both XGBoost models. Based on the parametric importance analysis conducted by the trained XGBoost models, a new universal correlation is developed by incorporating several key parameters to characterize condensation heat transfer. This correlation shows reliable predictions for all data points in the consolidated database with the mean relative deviation (MRD) of −1.77% and MARD of 19.21%. An additional excluded database is collected to validate the universality of the new correlation and these ML models. It illustrates that ML models are effective predictive tools for HTCs, and can also help develop the heat transfer correlation with high accuracy and universality.