We present a multi-modal deep learning framework for spacecraft orbit prediction that fuses historical trajectory data with real-time solar wind measurements via cross-attention. Using three years (2023–2025) of NASA SSC position data for ISS, DSCOVR, and MMS-1 combined with OMNI solar wind parameters, we train and compare bidirectional LSTM, Transformer, and residual gated fusion architectures for 6-hour trajectory prediction. Our LSTM achieves 125 km MAE on ISS at the 6-hour horizon, while the multi-modal architecture improves to 135 km during geomagnetic storms — a 17% improvement that validates solar wind as a meaningful leading indicator for LEO drag perturbations. We introduce a two-phase training strategy and sigmoid gating mechanism that guarantees the multi-modal model cannot underperform its single-modality baseline. All data, code, and trained model checkpoints are publicly available.
Ted Rubin (Sun,) studied this question.