Comparison between conventional and deep learning-based surrogate models in predicting convective heat transfer performance of U-bend channels

Abstract

Deep neural networks are efficient methods to achieve real-time visualization of physics fields. The main concerns that prevented deep learning from being implemented in the field of energy conversion were the risks of overfitting and the lack of data. Therefore, it is necessary to evaluate different kinds of surrogate modeling methods and provide guidelines for designers to choose models. In this study, three conventional models (Artificial Neural Network, Radial Bias Function, and Kriging), and two deep learning-based models (Convolutional Neural Network and Conditional Generative Adversarial Neural Network) were established to predict the flow and heat transfer performance of a U-bend with variable geometries. The models were detailly compared in terms of the single-point prediction accuracy, response accuracy, sensitivity to sample size, and other characteristics of interest. Results showed that the conventional models had slightly higher single point accuracy and the relative error of pressure loss and heat transfer were within ±6.6% and ±5.7% respectively, while those of the deep learning-based models were within ±8.0% and ±6.3% respectively. Nevertheless, the deep learning-based models had higher response accuracy and could reconstruct the distributions of surface pressure and wall heat flux with the pixel-wise absolute error within ±2.0 Pa and ±45 W/m2 respectively. The results indicated that deep learning was a promising surrogate modeling approach due to its acceptable prediction error and ability to reconstruct physical fields. This effort was expected to serve as a guide for establishing more reliable data-driven surrogate models for energy conversion and heat transfer problems.