AI Challenge for Satellite Tracking and Orbit Resilience Modeling (STORM‐AI): Data Set, Design, and Results
Sergio Sanchez‐Hurtado, Haley E. Solera, William E. Parker, Mia Tian, Ruoxi Qian, Edenna Chen, Enrico M. Zucchelli, Rasyid Ridha, Dominic Bruno, Christopher Yeung, Justin Gmys, Binh Tran, Bhargav M. Joshi, Victor Rodriguez‐Fernandez, Jefferson Mitchell, Morgan Mitchell, Jonathan P. How, Giovanni Lavezzi, Richard LinaresAbstract
Precise thermospheric density forecasting is critical for mitigating satellite drag in Low Earth Orbit (LEO), but traditional empirical models such as MSIS and JB2008 can fail during geomagnetic storms. To evaluate whether AI can better capture this transient behavior, the 2025 MIT ARCLab Prize for AI Innovation in Space asked participants to forecast orbit‐averaged density 3 days ahead. The training set included 8,118 samples, each with initial orbital parameters, 60 days of contextual space weather indicators, GOES X‐ray flux, OMNI2 solar wind observations, and the corresponding 3‐day density target, plus 4,557 hidden samples reserved for private testing. Submissions were ranked on Codabench using the Orbital Density Root Mean Square Error (OD‐RMSE), a custom metric that emphasizes early‐horizon performance and is benchmarked against MSIS. The competition attracted 145 teams and 973 submissions. Team Bimasakti won with a bias‐corrected ensemble of LightGBM models trained on ratio and log‐ratio features derived from observed and NRLMSIS 2.1 densities, Team Millennial‐IUP placed second with a hybrid CNN‐LSTM‐GRU model over a 30‐day window with physics‐guided scaling, and Team Digantara achieved strong public leaderboard performance with a combined XGBoost and TSLANet approach. Overall, the top approach achieved a 50.8% RMSE reduction versus JB2008 and a 67.7% reduction versus MSIS, although during the May 2024 geomagnetic storm these improvements dropped to 6.1% and 34.7%, respectively.