Falling film heat transfer from sensible convection to boiling on horizontal tubes: Database construction and machine learning-based heat transfer prediction

Abstract

Accurate predictions of heat transfer coefficients are essential for falling film evaporators’ optimum design and heat transfer enhancement. However, the complex interplay of mechanisms and various influencing factors, poses challenges for existing prediction methods. In this study, three general consolidated databases based on comprehensive reviews of relevant published literature were constructed, following which, artificial neural network (ANN) models employing the backpropagation (BP) algorithm were developed to predict falling film heat transfer coefficients in sensible convection, evaporation, and boiling. The sensible convection database comprises 1635 data of water and alcoholic fluids, the evaporation database contains 2097 data of water, R22, and ammonia, and the boiling database consists of 1207 data involving water, R134a, R123, R11, R32, R290, R600a, R1234ze, and R245fa. An optimal ANN model was selected for prediction of the test dataset after evaluating impacts of neural network architectures and parameter combinations on the model performance. The mean absolute errors (MAEs) of three databases are 1.29 %, 1.09 %, and 5.02 %, respectively, with coefficient of determination (R²) of 0.9177, 0.9236, and 0.9637, respectively. Sensitivity analyses of the input parameters were performed to strengthen the interpretability of the model's construction. The superiority of the present optimal ANN model was revealed through comparisons with published correlations. Furthermore, the prediction capability of the present model for entirely new and unfamiliar data was assessed, demonstrating its substantial generalization capability.