A Graph Neural Network Approach for High-Fidelity Thermospheric State Estimation

Accurate satellite orbit prediction in low-Earth orbit (LEO) has become increasingly important as satellite congestion grows in this region of the atmosphere, increasing the risk of collisions with other spacecraft and space debris. Atmospheric drag is the dominant source of uncertainty at LEO altitudes, with it being the largest non-conservative force in this region. This makes accurate estimation of thermospheric parameters essential for reliable orbit propagation, as the LEO drag force is a function of neutral thermospheric parameters.

Orbit prediction relies on estimating thermospheric properties such as density, temperature, and winds, where currently empirical or numerical models are used to generate these values. Although widely used in space operations, these models struggle to capture the thermosphere’s dynamic behaviour, which leads to significant errors in drag estimation and orbital predictions. A major event occurred in February 2022, when SpaceX launched 49 Starlink satellites during a minor geomagnetic storm. This unexpectedly increased satellite drag, causing 38 satellites to deorbit, where they were ultimately lost as a result of atmospheric reentry. This loss for SpaceX shed light on the need for more accurate thermospheric models, as satellite operators rely heavily on these models for orbit planning.

As collision risk increases, satellite operators require higher-fidelity modelling approaches. While machine learning methods have shown promise in improving thermospheric state prediction, they are not yet widely adopted. Graph Neural Networks (GNNs) have demonstrated strong performance in spatiotemporal modelling of complex geophysical systems. Notably, Google DeepMind’s GraphCast model demonstrated predictive skill comparable to that of the ECMWF operational forecasting system, setting a new benchmark for medium-range tropospheric weather prediction.

This research develops a GNN-based framework to model the spatiotemporal dynamics of the thermosphere, enabling improved estimation of neutral atmospheric parameters and supporting more accurate orbit prediction in the near-Earth environment.