System-wide spacecraft temperature estimation using graph neural networks

• A Graph Neural Network–based model was proposed for spacecraft thermal analysis. • Thermal mathematical models were represented as graph data to enable GNN learning. • The model accurately estimated system-wide temperatures in large-scale spacecraft. • Achieved <1 K error with less than half of the computational cost of previous method. During a mission, the thermal state of a spacecraft can change because of factors such as variations in component contact conditions and degradation in the surface optical properties. These uncertainties make it difficult for the pre-launch thermal mathematical model (TMM) to estimate the on-orbit temperature distribution accurately. To address this issue, various methods have been developed over the years to estimate the on-orbit temperature distribution using partial onboard temperature measurements. However, conventional methods struggle to predict the temperature distribution of large-scale spacecraft systems with >1000 nodes. In this study, we proposed a method based on graph neural network (GNN) to incorporate thermal coupling structures into an inference model, enabling efficient estimation of the temperature distributions in large-scale spacecraft systems. The complex thermal structure of the spacecraft was represented as graph data composed of nodes and edges to which GNN was applied. Numerical experiments demonstrated that by learning based on the thermal structure of the spacecraft, the proposed method achieved large-scale temperature prediction with an average error of about 1 K, which was less than half of the computational cost of previous methods. The results indicated that GNN could integrate the structural features inherent in spacecraft thermal phenomena into the learning process, thereby allowing flexible scalability for large systems. Thus, the proposed method represents spacecraft TMMs that integrate physical insights with machine learning.