Speaker
Description
Spacecraft thermal analysis relies on high-fidelity numerical models, but repeated transient simulations across different orbital and operational conditions can be computationally expensive. This work presents a physics-guided modular neural surrogate developed using ESATAN-TMS to accelerate system-level transient thermal predictions.
The reference spacecraft model comprises 204 lumped thermal nodes organised into 16 physical subsystems. Environmental heat loads are estimated by two dedicated neural models. The first predicts direct solar radiation and planetary albedo outside eclipse, while a deterministic physics-based gate sets both contributions to zero during eclipse. The eclipse detector achieved complete agreement with ESATAN-TMS across all evaluated cases. The second model predicts planetary infrared radiation throughout the entire orbit, as this contribution remains present during eclipse. Temperature evolution is then predicted by a Stable Block Graph GRU combining local subsystem encoders, inter-block attention and a stable update formulation designed to limit autoregressive error accumulation.
The dataset comprises 10,000 ESATAN-TMS cases, each containing 1,001 time samples with a 10 s timestep and covering different orbital, attitude and internal-dissipation conditions. The cases were divided into non-overlapping training, validation and test sets using a 70%/15%/15% split. On the test set, the solar plus albedo and planetary infrared models achieved mean absolute errors of 0.0521 W and 0.0141 W, corresponding to 1.06% and 1.18% of their respective mean heat loads. The thermal surrogate achieved a global MAE of 0.0991 K over 997 consecutive free-run prediction steps. The mean node-wise MAE corresponded to only 0.225% of the nodal temperature excursion. The maximum absolute temperature error was 1.607 K, while the maximum sustained relative error was 2.94%. At the final prediction step, the MAE remained limited to 0.1184 K, confirming that error accumulation remained contained throughout the transient sequence.
Once trained, the surrogate evaluates a complete 10,000 s transient scenario in approximately 4 s, compared with 53–59 s for a single ESATAN-TMS simulation in the tested configuration. The proposed framework therefore enables rapid exploration of operating conditions and may also support future transfer-learning-assisted updating and correlation of high-fidelity thermal models.