Physics-informed deep residual neural network for finned-tube evaporator performance prediction

Abstract

Finned-tube heat exchangers are extensively used in industrial and daily life. As complex system designs and digital twins increase, higher demands arise for predictive modeling of finned-tube heat exchanger performance. Previous studies faced issues like disparities in evaporator performance under dry and wet conditions, inaccurate predictions of refrigerant pressure drop, and a lack of comprehensive consideration for predicting the trends in physical details. With the evolution of AI algorithms, exploring novel modeling methods becomes essential. Therefore, this study proposes a physics-informed deep residual network model for accurate performance prediction of finned-tube evaporators. The model achieved mean absolute error of 0.14 % for heat transfer rate prediction, 0.36 % for refrigerant pressure drop prediction, and 0.49 % for air pressure drop prediction. This model addresses previous issues of performance disparities under dry and wet conditions and high errors in refrigerant pressure drop prediction. The proposed deep residual network model is found much better performance in capturing physical details than other machine learning approaches. It can serve as a benchmark model for transfer learning or complex system modeling. The method of determining the structure of residual networks through a priori physical knowledge also provides a new perspective on modeling physical systems.