Surrogate modeling for spacecraft thermophysical models using deep learning

Thermal modeling is a critical technology in spacecraft thermal control systems, where the complex spatially and temporally variable parameters used as inputs to the spacecraft usually result in long operation times, which hinders sensitivity analysis and control strategies. The uniqueness of both the space environment and the working conditions of each spacecraft, make thermal models differ in different space environments; thus, traditional thermal modeling methods that rely heavily on physical knowledge need to build more than one corresponding thermal model, and they also cannot generalize well. Therefore, an intelligent surrogate modeling strategy for spacecraft thermophysical models that uses deep learning, called SMS-DL, is proposed. An intelligent batch processing system for thermal analysis based on NX TMG Thermal Analysis was designed to automate the input of the thermal design parameters and the output of the thermal analysis results through macro-recording and playback using a state-of-the-art biconjugate gradient solver to provide superior speed, reliability, and accuracy, thus achieving a trade-off between high accuracy and low computational cost. One deep neural network based on Bayesian optimization was pre-trained using the thermal analysis data of the spacecraft in the source domain calculated using a batch processing system, which had a computational speed that was 1000+ times faster than that of the traditional thermophysical model and high computational accuracy of 99%+. Then, it was applied to the target domain with a limited amount of thermal analysis data using model fine-tuning. The theoretical and experimental results from the thermal analysis modeling of the near-ultraviolet radiation detector on the China Space Station Telescope developed in China demonstrated that deep transfer learning effectively adapted the pre-trained model from one working condition to another, improved the prediction accuracy by at least 86.4% over the direct prediction accuracy using the pre-trained model, and had better predictive performance than learning from scratch with only a limited amount of data.