Unified physics-informed ResNet model for comprehensive performance prediction of finned-tube evaporators.

Finned-tube heat exchangers are widely used in refrigeration and thermal engineering. With the growing demand for digital twins in complex systems, the need for fast and accurate performance prediction has become more urgent. Existing neural network models often suffer from fragmented structures, incomplete parameter considerations,
and an inability to capture essential physical details. This study presents a unified physics-informed residual network model that overcomes these limitations by ensuring parameter completeness and integrating physical principles into the network architecture. The proposed method avoids redundant input parameters and conforms to physical laws. Moreover, it ensures that the selected output parameters meet the requirements of actual performance prediction and are conducive to the training of neural networks. By combining residual blocks with independent layers, the model achieves joint prediction of multiple performance metrics, improving both efficiency and accuracy. The proposed model significantly enhances predictive performance. All the determination coefficients reach 0.999, while both the mean absolute error and root mean square error values
remain remarkably low. Specifically, it achieves a mean absolute percentage error of 0.41 % for total heat transfer rate, 0.65 % for sensible heat transfer rate, and below 0.3 % for both refrigerant and air pressure drop. Furthermore, it effectively captures critical physical details, such as transitions between dry and wet operating conditions. Compared to previous models, this approach provides a more comprehensive, physically consistent, and computationally efficient framework for finned-tube evaporator performance prediction.