Machine Learning–Based Correlation Methods for Satellite Thermal Analysis Models

Accurate thermal analysis is essential for reliable satellite thermal design. To improve the accuracy of thermal analysis, thermal analysis model correlation is performed using thermal vacuum chamber test results. However, conventional correlation methods, dependent on manual trial-and-error or optimization algorithms, face challenges in terms of both accuracy and efficiency when dealing with large numbers of analysis parameters. To address these limitations, this study proposes two machine learning–based correlation methods. The first approach utilizes machine learning to construct a neural network for a surrogate model, applying an optimization algorithm to estimate analysis parameters for correlation. The second approach utilizes machine learning by training a neural network to directly estimate analysis parameters from the target temperature, eliminating the need for optimization. According to performance evaluation results using a CubeSat model, both proposed correlation methods demonstrated superior performance with shallow neural networks trained using the Levenberg–Marquardt algorithm and outperformed existing techniques in terms of accuracy and efficiency. The findings from this study can be applied directly to correlate satellite thermal analysis models with experimental data from thermal vacuum chamber tests, resulting in improved correlation accuracy.