Abstract: A neural network-based inversion method for efficient calculation of structural deformation of spacecraft in orbit was proposed in this paper. Finite element analysis was used to obtain the deformation distribution characteristics of the structure under different temperature loads. The achieved data was applied to train the neural network between input and output to get a high-precision proxy model. By this model, the temperature measured in orbit was applied as input to invert the full field deformation of spacecraft structures rapidly. The adaptability of neural networks to input errors was enhanced by adding Gaussian noise with appropriate parameters. Using an improved weight connection analysis method, the layout optimization of the temperature measurement points of the structure with the highest deformation inversion accuracy was given for determining the number of sensors. The proposed method featured by high accuracy, strong real-time performance, and minimal susceptibility to input errors, is of significance for improving the imaging quality of remote sensing satellites.