Data‐Driven Satellite Drag Modeling Without Dynamic Knowledge of the Atmosphere

Abstract Increasing congestion in low Earth orbit (LEO) from new satellite deployments and a growing debris population has made accurate orbital trajectory prediction essential for anticipating and avoiding potential collisions on a regular basis. Atmospheric drag is usually the dominant source of propagation error across much of LEO, especially during geomagnetic storms when large portions of the satellite catalog often become poorly tracked for days. The conventional drag modeling approach typically relies on thermospheric neutral density and satellite ballistic coefficients, two quantities that are very difficult to know or measure in practice. In addition to the data scarcity issue, the conventional drag modeling approach imposes rigid structure and high computational cost, which makes uncertainty estimation difficult and increases latency. This work introduces REACT: Response Estimation and Analysis for Correlated Trajectories, a new satellite drag modeling framework that exploits correlated drag responses to space weather across satellites in LEO. REACT captures both the direct mapping from space weather drivers to observed satellite dynamics and the shared correlations among tracked satellites. The end‐to‐end, data‐driven framework avoids reliance on dynamic neutral density models or prior knowledge of satellite ballistic coefficients. Gaussian conditioning is used to infer satellite drag responses by leveraging historical relationships among measured quantities. This approach means that any measurement of any space weather driver or satellite state could be used to update our belief in satellite states across the entire catalog in near‐real time. REACT offers improved satellite tracking robustness in LEO by offering a fast, flexible drag model that captures uncertainty and reduces latency while requiring no prior knowledge of atmosphere dynamics or satellite properties.