Deep Learning Based Thermal Stress and Deformation Analysis of Satellites

Abstract

When analyzing the thermal stress and deformation of satellites in orbit, the traditional numerical methods, such as the finite difference and the finite element, are expensive and time-consuming. To improve computational efficiency, we propose a deep-learning based surrogate to immediately predict the thermal stress and deformation of a satellite with a given temperature field, where the U-Net is employed to learn the end-to-end mapping from the temperature field to the thermal stress and deformation. A data set with less smooth temperature fields is generated to augment the training data, by which the accuracy and generalization performance of the model is significantly improved. Combined with a rapid temperature prediction method, the model predicts the thermal stress and deformation of a satellite motherboard given several heat sources, verifying the feasibility and effectiveness of the proposed method.