Machine learning-based models for frictional pressure drop prediction of condensing and adiabatic flow in micro, mini and macro channels utilizing universal data

This study proposes a universal machine learning-based model to predict the adiabatic and condensing frictional pressure drop. For developing the proposed model, 11,411 data points of adiabatic and condensing flow inside micro, mini and macro channels are collected from 80 sources. The database consists of 24 working fluids, hydraulic diameters from 0.07 to 18 mm, mass velocities from 6.3 to 2000 Kg/m2s, and reduced pressures from 0.001 to 0.95. Using this database, four machine learning regression models, including “artificial neural network”, “support vector regression”, “gradient boosted regression”, and “random forest regression”, are developed and compared with each other. A wide range of dimensionless parameters as features, “two-phase friction factor” and “Chisholm parameter” are each considered separately as targets. Using search methods, the optimal values of important hyperparameters in each model are determined. The results showed that the “gradient boosted regression” model performs better than other models and predicts the frictional pressure drop with a mean absolute relative deviation of 3.24%. Examining the effectiveness of the new model showed that it predicts data with uniform accuracy over a vast range of variations of each flow parameter.