Attenuation Field Prediction Using Physics-Guided Spatio-Temporal Diffusion in LEO Satellite-to-Ground Communications

Abstract

Accurate attenuation field prediction for low earth orbit satellite-to-ground (StG) is critical for capturing the dynamic variations of signal propagation under complex atmospheric and geometric conditions. Existing StG attenuation modeling implicitly assumes attenuation is a point-wise scalar, which is insufficient for capturing such field-level variations. In this work, we propose PG-STD, a physics-guided spatio-temporal diffusion model that generates physically consistent attenuation fields. PG-STD unifies ITU-R physics constraints with multimodal conditioning on satellite and atmospheric observations, enabling physically consistent and channel-adaptive attenuation field generation. A physics-informed learning objective further improves interpretability and link reliability. We construct a large-scale radar-satellite dataset across diverse atmospheric conditions with Starlink orbital information, and additionally validate our model on a public NASA weather dataset. Experimental results show that PG-STD achieves higher attenuation structure correlation and lower physics consistency errors than baselines, demonstrating its effectiveness in generating accurate and interpretable StG attenuation fields. To evaluate system-level applicability, the PG-STD is further integrated into a satellite power allocation algorithm, demonstrating that it enables the allocation algorithm to respond more effectively to dynamic channel variations.