Speaker
Description
Accurate knowledge of the Earth's thermosphere is vital for a range of space operations, including space traffic management, collision avoidance, re-entry predictions, and orbital lifetime analysis. As Low Earth Orbit (LEO) becomes increasingly congested, precise orbit determination and highly accurate conjunction assessments are of critical importance.
The primary source of uncertainty in LEO trajectory propagation stems from limitations in existing empirical thermospheric density models, such as JB2006, JB2008, and NRLMSISE-00. Because aerodynamic drag is directly proportional to atmospheric density ($\rho$) and a satellite's drag coefficient (often assumed static at $C_d \approx 2.2$), density estimation errors cascade directly into positional uncertainties. While empirical models perform adequately during quiescent periods, extreme space weather phenomena introduce high levels of uncertainty, as these static models struggle to capture highly non-linear temporal atmospheric dependencies.
To address this, this talk presents a systematic evaluation of existing empirical baselines alongside a novel machine learning-based calibration framework designed to mitigate these systematic modelling errors. We propose a Recurrent Time Delay Neural Network (RTDNN) architecture that combines predicted densities with high fidelity measurements obtained from existing satellite instrumentation (CHAMP, GRACE, and Swarm). By explicitly passing time shifted solar and geomagnetic indices as input features, this autoregressive approach successfully captures the thermosphere's thermal inertia and delayed response to solar forcing.
The error reductions presented in this talk span diverse solar regimes, including deep minimums, moderate fluctuations, and extreme geomagnetic storms. Results demonstrate a drastic reduction in Mean Absolute Percentage Error (MAPE) compared to standalone empirical models. Notably, during the deep solar minimum of 2009, the calibrator reduced the NRLMSISE-00 error from 65.3% to 0.86%, and during the extreme G5 solar storms of May 2024, it corrected the JB2006 model to achieve an error rate of just 0.21%.
By significantly reducing density estimation errors without sacrificing computational efficiency, this ML calibrated framework prevents the artificial inflation of orbital drag forces. Implementing this architecture into operational flight dynamics systems will yield more accurate Collision Data Messages (CDMs), reduce false alarm avoidance manoeuvres, and directly support ESA's Zero Debris objectives. Finally, this talk will outline our ongoing research roadmap, which explores the application of Liquid Neural Networks (LNNs) to further enhance the framework's real time adaptability to continuous, non stationary space weather dynamics.