Physics-Gated Neural Networks for Resilient Satellite Drag Correction During Geomagnetic Storms

The Simplified General Perturbations-4 (SGP4) propagator is widely used for satellite tracking but suffers accuracy degradation during geomagnetic storms due to unmodeled atmospheric drag variations. This work introduces a Gated Physics-Informed Neural Network (G-PINN) that applies a differentiable physics gate to enforce zero error at the initial epoch while learning non-linear drag corrections over time. The model integrates orbital state vectors with space weather indices such as F10.7 solar flux and is trained on historical Starlink trajectory data (2022–2025). Experimental results show a 12.6% reduction in mean positional error during severe geomagnetic storms while maintaining stability during quiet conditions. Additionally, the model identifies a systematic radial velocity bias in SGP4-based TLE data, demonstrating its capability as both a correction and diagnostic system. A Monte Carlo Mahalanobis-distance-based conjunction assessment framework is also implemented to evaluate collision risk.