Physics-informed neural networks for high-fidelity thermal modeling of space infrared payloads

Abstract

Precise and efficient thermal management technology serves as a critical support for the normal operation of space infrared payloads, while high-fidelity thermal models constitute the foundation for conducting various thermal analyses. Traditional numerical modeling methods rely on mesh discretization and iterative solvers; on one hand, this introduces geometric errors and entails a substantial preprocessing workload for the payload's intricate components; on the other hand, the strong nonlinearity of thermal radiation and the time-consuming nature of iterative computations make it difficult to satisfy the requirements for real-time, on-orbit thermal analysis. This paper proposes a high-fidelity modeling framework for space infrared payloads based on Physics-Informed Neural Networks (PINNs). The payload's geometric information is input into the PINN as coordinate data, enabling a naturally mesh-free solution within a continuous domain. Furthermore, thermal conduction equations and radiative boundary conditions are incorporated as physical constraints for the PINN, ensuring physical consistency and enabling rapid inference after training. Within this framework, the PINN solves the forward problem to predict the full-field temperature distribution under on-orbit thermal environments. The Optimus optimization platform invokes the PINN to solve the inverse problem, identifying key thermal parameters such as surface emissivity. The combination of forward-problem prediction and residual-driven parameter correction establishes a closed-loop modeling process, ultimately achieving the construction of a high-fidelity thermal model. Validation against NX TMG reference solutions demonstrates a maximum relative error of 0.06% and a mean absolute temperature deviation of 0.17 K, demonstrating the high accuracy of the proposed framework for space infrared payload thermal modeling. Once the PINN training is complete, obtaining the full-field temperature distribution requires only a matter of seconds for inference, thereby providing an effective technical pathway toward realizing intelligent thermal management for space-based infrared payloads.