Thermal analysis is critical to spacecraft mission success. However, uncertain parameters such as contact conductance and effective emissivity make it difficult to predict temperatures accurately. Therefore, uncertainty quantification (UQ) is important in the thermal analysis of spacecraft. In recent years, Monte Carlo simulation (MCS) has been used as one of the most useful methods for UQ. On the other hand, Monte Carlo simulations are computationally expensive because they require a large number of physical simulations. To significantly reduce computational cost, a surrogate model with data-driven has been proposed. However, many have limited temperature prediction points and do not predict system-wide temperatures. If system-wide temperatures could be predicted at low cost, a different approach to thermal analysis could be taken. This study proposes a transient surrogate model using proper orthogonal decomposition (POD) and Long Short-Term Memory (LSTM), a type of recurrent neural network that can predict the entire system and sequence at a low cost. The spatial mode U is extracted from the prior data set by POD, and the coefficients ? of the spatial mode U are predicted each time by LSTM. The use of POD allows for low-cost system-wide prediction, while the use of LSTM allows for prediction against time-varying uncertainty. In this study, the method was applied to the satellite thermal model. Numerical experiments were conducted on a satellite thermal model to compare the accuracy and analysis time of surrogate and thermal mathematical models. The numerical experiments show that the proposed method can significantly reduce the analysis time without losing accuracy.
Yamashita et al. (Sun,) studied this question.