Physics-embedded neural-network surrogate for infrared radiation prediction of correlated space targets

Infrared radiation characteristics are key electromagnetic signatures for detecting, tracking, and assessing space targets. This study develops a physics-embedded neural-network surrogate for rapid prediction of the infrared radiative intensity of correlated targets with different surface emissivities. The model uses the radiation time series of two basic targets as inputs and infers two effective heat-flux sequences: a full-band external heat flux governing temperature evolution and a detection-band ambient projected heat flux governing reflected radiation. The heat balance equation and the Planck-law radiation calculation are embedded as forward computational steps, and the model is trained by a weighted RMSE between reconstructed and reference radiation intensities of the two basic targets. It is a physics-embedded surrogate in which the governing equations constrain the input-output structure. Numerical tests are conducted within a synthetic simulation chain using Aerospace ToolKit trajectories and an in-house MATLAB thermal-radiation model. Over the emissivity range 0.05-0.95, the long-wave infrared RRMSE is generally below 2%, while the mid-wave infrared RRMSE remains below 5% for emissivity ≥ 0.2 and reaches 11.55% at emissivity = 0.05. The main errors occur for low-emissivity mid-wave cases and during eclipse-to-sunlight transitions. The results show that hard embedding of radiation physics can reproduce simulator-generated emissivity trends with interpretable effective latent variables.