Physics-Informed Machine Learning Towards A Real-Time Spacecraft Thermal Simulator

Modeling thermal states for complex space missions, such as the surface exploration of airless bodies, requires high computation, whether used in ground-based analysis for spacecraft design or during onboard reasoning for autonomous operations. For example, a finite-element-method (FEM) thermal model with hundreds of elements can take significant time to simulate on a typical workstation, which makes it unsuitable for onboard reasoning during time-sensitive scenarios such as descent and landing, proximity operations, or in-space assembly. Further, the lack of fast and accurate thermal modeling drives thermal designs to be more conservative and leads to spacecraft with larger mass and higher power budgets. The emerging paradigm of physics-informed machine learning (PIML) presents a class of hybrid modeling architectures that address this challenge by combining simplified physics models (e.g., analytical, reduced-order, and coarse mesh models) with sample-based machine learning (ML) models (e.g., deep neural networks and Gaussian processes) resulting in models which maintain both interpretability and robustness. Such techniques enable designs with reduced mass and power through onboard thermal-state estimation and control and may lead to improved onboard handling of off-nominal states, including unplanned down-time (e.g. GOES-7 [1]). The PIML model or hybrid model presented here consists of a neural network which predicts reduced nodalizations (distribution and size of coarse finite difference mesh) given on-orbit thermal load conditions, and subsequently a (relatively coarse) finite-difference model operates on this mesh to predict thermal states. We compare the computational performance and accuracy of the hybrid model to a purely data-driven neural net model, and a high-fidelity finite-difference model (on a fine mesh) of a prototype Earth-orbiting small spacecraft. The PIML based active nodalization approach provides significantly better generalization than the neural net model and coarse mesh model, while reducing computing cost by up to 1.7 x compared to the high-fidelity model.