Physics-informed neural network for rapid prediction of the temperature fields in fuel-heat pipe assemblies

Accurate prediction of the temperature distribution within the fuel-heat pipe assemblies is crucial for the design optimization and safety analysis of heat pipe reactors. A Physics-Informed Neural Network (PINN) model is proposed for rapid prediction of the temperature fields in fuel-heat pipe assemblies under varying boundary conditions. The effects of various loss functions (MSE, MAE, Huber Loss) and sampling point numbers on the model performance were investigated. Among these, the MSE demonstrates the best performance, achieving an average relative error of 0.114 %. Furthermore, increasing the sampling points can improve the prediction accuracy, with the training time increasing sublinearly due to parallel computation. In order to improve the model performance, three model types (purely data-driven, purely physics-driven, and hybrid-driven) were analyzed. The results showed that the average relative error of the hybrid-driven model is the lowest at 0.09763 %. In order to improve the model prediction performance and training speed under target domain shifts, transfer learning method was introduced. For the purely data-driven model trained with only ten cases, transfer learning reduces the average relative error from 8.031 % to 0.4107 %. The hybrid-driven model also benefits from transfer learning, decreasing the average relative error from 0.1179 % to 0.08879 %. This study demonstrates the effectiveness of the PINN in temperature field prediction for engineering applications.