Artificial Intelligence for Spacecraft Thermal Control Systems

The proposed innovation involves the creation of a user-friendly software tool for the generation of surrogate models for thermal modeling using Physics-Informed Neural Networks (PINNs) and Physics-Informed Deep Operator Networks (PI-DeepONets). When users provide basic system parameters (e.g., geometry, material properties, nominal boundary conditions), the software will generate the relevant physics-based loss functions and boundary/initial conditions for PINN or DeepONet models. Under the hood, adaptive training schedules adjust the relative weighting of physics losses and data-driven losses, ensuring consistent convergence across a wide parameter space. Users will be provided feedback on training progress, residual errors, and parameter coverage, helping engineers quickly diagnose convergence issues or refine the problem setup. We will also investigate the reliability of generating thermal results directly from CAD geometry, which would significantly reduce the human labor involved with thermal modeling by avoiding the need to mesh geometry. With this automated pipeline, spacecraft thermal engineers can rapidly explore numerous configurations of spacecraft designs without the overhead of manually re-deriving equations or building new surrogate models from scratch. We will use our validated heat transfer software TAITherm/MuSES to generate training and validation data for the physics-informed surrogate models. Once trained, the surrogate models generate reliable temperature predictions in near- real-time, enabling accelerated design iterations and on-demand thermal analysis for operational missions. The tool will target NASA and aerospace industry markets, including commercial and government spacecraft engineers seeking streamlined, cost-effective, and reliable thermal modeling solutions.