Deep learning meets orbital mechanics. We trained LSTM, Transformer, and multi-modal models on 3 years of NASA spacecraft data to predict orbits and measure the impact of solar wind on trajectory perturbations.
International Space Station. Low Earth Orbit with significant atmospheric drag. 1.58M data points. Most affected by thermosphere expansion during geomagnetic storms.
Deep Space Climate Observatory at the Sun-Earth L1 Lagrange point. 131K data points. Directly measures solar wind before it reaches Earth.
Magnetospheric Multiscale Mission. Highly elliptical orbit cutting through Earth's magnetosphere. 1.54M data points. Most sensitive to space weather conditions.
Encoder-decoder LSTM with autoregressive decoding. Maps 24h of position/velocity history to 6-24h predicted trajectories.
PyTorch Sequence-to-SequenceEncoder-decoder Transformer with sinusoidal positional encoding and learned query tokens. Multi-head cross-attention.
PyTorch AttentionDual-encoder combining orbit positions with solar wind measurements. Cross-attention lets the model learn which solar wind conditions most affect the trajectory.
Cross-Attention Solar WindPhysics-based orbit propagation using two-body Keplerian mechanics. The benchmark to beat. ML models target short-horizon improvement.
Baseline Physics6-hour prediction MAE (Mean Absolute Error) in km. Trained on dual RTX 5090 GPUs via RunPod.
| Model | ISS (LEO) | DSCOVR (L1) | MMS-1 (HEO) |
|---|
Real-time International Space Station position on a 3D globe. Updated every 5 seconds.
Current solar wind conditions from NOAA SWPC. These are the signals our multi-modal model uses to predict orbit perturbations.
Run orbit predictions directly in your browser using our trained models on Hugging Face.
Automated daily fetch of spacecraft positions (GSE/GEO) and solar wind parameters (IMF, flow speed, density, Kp, Dst) via cron.
Velocity derivation, normalization, gap detection, sliding window creation. L1-to-Earth propagation delay correction for solar wind.
AdamW + cosine LR schedule, early stopping, gradient clipping. Temporal train/val/test split (70/15/15). Both PyTorch and TensorFlow.
MAE/RMSE in km at 1h, 6h, 24h horizons. Separate analysis for quiet (Kp ≤ 3) vs storm (Kp > 5) conditions.
We hypothesize that during geomagnetic storms, the thermosphere heats and expands, increasing atmospheric drag on LEO satellites in ways that SGP4 cannot predict. By incorporating real-time solar wind data measured at L1 (~45 min before arrival at Earth), our multi-modal model can anticipate these perturbations. The dataset captures Dst down to -406 nT and Kp up to 8+ — severe storm events ideal for testing this hypothesis.