Surrogate modeling for physical fields of heat transfer processes based on physics-informed neural network

By constructing structured deep neural network architecture, physics-informed neural networks (PINN) can be trained to solve supervised learning tasks with limited amount of boundary data while effectively integrating any given laws of physics described by general nonlinear partial differential equations (i.e., Navier-Stokes equation). However, the effect of PINN training is closely related to how the boundary conditions are set. In this work, two 2-D steady-state heat transfer problems, namely heat conduction model with internal heat source and convection heat transfer equation between plates are taken as examples. Two surrogate models are trained based on PINN by using two setting methods of soft boundary and hard boundary. The trained surrogate models are used to predict the output of temperature fields, which are verified and compared with the simulated data. The comparison results show that the prediction ability of PINN based on hard boundary is superior to the rival. © 2021, Editorial Board of CIESC Journal. All right reserved.