Physics-Informed Neural Network for Thermal Analysis of Space Structure

Abstract

External heat flux often induces deformation or vibration in space structures comprised of thin-walled tubes. Efficient and real-time thermal-structural dynamic analysis is essential for the reliable operation and optimal design of such a structure. However, traditional finite element method (FEM) requires a significant amount of time for thermal-structural dynamic analysis of the complex space structure. The present work proposes a multi-boundary condition physics-informed neural network (mb-PINN) to address the thermal governing equation (TGE) of thin-walled tubes under different incident angles of heat flux. Specifically, the proposed mb-PINN constructs an independent neural network to fit the mapping relationship between incident angle and boundary condition. The variable boundary condition is then integrated into PINN by taking the incident angle as a feature input of PINN. Moreover, a dynamic sampling method is further incorporated into mb-PINN, which reallocates sampling points to improve the accuracy of the approximate solution of TGE. The thermal structural behavior of the tube under different incident angles of heat flux can therefore be predicted quickly and accurately, offering an efficient solution for the thermal-structural dynamic analysis of space structure.