Efficient Thermal Analysis of On-Orbit Satellites Using a Nondimensionalized Physics-Informed Neural Network

This study proposes a nondimensional physics-informed neural network (Nd-PINN) framework for predicting the infrared radiation characteristics of low-Earth-orbit satellites under complex space thermal environments. The Laplace heat conduction equation and nonlinear boundary conditions involving solar irradiation, Earth albedo, and Earth infrared radiation are embedded into the loss function, while nondimensional preprocessing improves training stability by balancing residual magnitudes. Four numerical cases of increasing complexity were used for validation. For cases with analytical solutions, Nd-PINN achieved absolute temperature errors below 0.2 K, relative infrared radiation errors below 0.3%, and energy conservation errors below 0.001%. Under unidirectional solar irradiation, the model shows good consistency with the azimuthal temperature gradient predicted by Lambert’s cosine law, with an energy conservation error below 0.58%. In the multi-source coupled radiation case, it maintained strong physical consistency, with an energy conservation error below 0.54%. These results demonstrate that Nd-PINN can effectively solve both linear and strongly nonlinear thermal problems without complex meshing, while providing continuous and smooth global temperature fields. The proposed framework offers an accurate, mesh-free alternative for satellite thermal analysis and infrared feature identification.