Adaptive AI Forecasting of Thermospheric Neutral Density Tuned to GRACE Data

Accurate forecasting of Thermospheric Neutral Density (TND) is essential for Low-Earth Orbit (LEO) mission planning, collision avoidance, and orbit determination. Atmospheric drag strongly influences satellite trajectories below 1000 km altitude, making precise density estimates critical for operational safety. Current empirical and physics-based modelsoften show limited skills to capture short-term variability driven by solar and geomagnetic activity. This limitation reduces their accuracy during dynamic space weather conditions and impacts mission planning.

We propose an adaptive machine learning framework using Extreme Gradient Boosting (XGBoost) to predict the systematic deviations from NRLMSISE-2.1 in log space. The model combines GRACE accelerometer-derived TND observations for the years 2009-2017, CODE's global TEC maps, and space weather indices represented by indices such as F10.7 and Ap. Feature engineering incorporates diurnal and seasonal cycles, altitude dependence, and ionosphere-thermosphere coupling. We apply lagged TEC and geomagnetic indices for short-term memory without needing sequential models. This ensures that this approach stays compatible with tabular workflows and keeps them computationally efficient.

A warm-start learning scheme is introduced tofacilitate short-term adaptation through fine-tuning the model with respect to the most recent observations. Validation on the GRACE and Swarm datasets shows an improvement compared to the original NRLMSISE-2.1 model. The reduction in RMSE is approximately 60-70%, and a MAPE improvement of a similar margin is seen under quiet conditions. Storm-time robustness has also been improved. The model performs well when validated on an off-track manner to validate its spatial generalization properties beyond the nominal orbit covered by the GRACE mission. The RMSE reduction is approximately 40%,

These results highlight the potential of AI-driven approaches for operational thermospheric density forecasting. Improved accuracy supports orbit prediction, drag estimation, and space weather applications. The novel framework combines robustness, adaptability, and computational efficiency. This makes it appropriate for integration into real-time mission planning and collision avoidance systems.