MIT ARCLab STORM-AI Competition
- Atilla Saadat
- May 30
- 3 min read
Updated: 3 hours ago
Real‑Time Thermospheric Density Forecasting with a BiGRU‑Attention Network
STORM‑AI Phase 2 Project Recap
Why I Tackled Thermospheric Density Forecasting:
Low‑Earth Orbit (LEO) is no longer roomy real estate: more than 5,500 operational spacecraft already circle Earth, and the mega‑constellations on the launch manifest will multiply that figure in the next few years . When the Sun hurls a geomagnetic storm our way, the thermosphere can swell by an order of magnitude, amplifying drag and scrambling conjunction assessments. Today’s options are unsatisfying:
First‑principles general‑circulation models (e.g., TIE‑GCM, GITM) track fine‑scale physics but demand hours on a super‑computer for a single 3‑day forecast .
Empirical climatologies (NRLMSISE‑00, JB2008) respond in milliseconds yet miss post‑storm cooling and routinely stray by > 40 % during disturbances .
I wanted a real‑time model that keeps the speed of climatologies but inherits the storm awareness of physics codes. The MIT ARCLab STORM‑AI Challenge supplied the perfect testbed: predict orbit‑averaged density at 10‑min cadence for the next 72 h, and be judged by an exponentially time‑weighted OD‑RMSE skill score. That precise metric, plus a hard runtime envelope, drove every design choice in my solution.
A 60‑Day History In, a 72‑Hour Forecast Out:
Data buffet (181 features × 1,440 time‑steps)
Each training sample is a 60‑day slice of hourly drivers:
168 dynamic OMNI2 channels – solar‑wind plasma, IMF components, geomagnetic indices, plus explicit lags at {1, 2, 3, 4, 6, 12} h and a 3 h rolling mean to expose multi‑scale variability .
Orbit context – six Keplerian elements and derived latitude/longitude/altitude.
Temporal encodings – sin/cos of longitude, day‑of‑year, and sidereal time.
Sequence‑wise normalizer ρ₀ removes altitude trends so the network learns relative fluctuations transferable across satellites .
A two‑stage scaler (quantile → standard) maps every feature to Ɲ(0, 1), taming proton‑flux outliers and ensuring numeric stability.
Stage | Layer | Output |
Encoder | 3 × BiGRU (h = 384/dir) | 1440 × 768 hidden states |
Attention | Additive attention | 768‑D context vector, z |
Head | LayerNorm → GELU → Linear | 432‑step log‑density residual |
The additive attention automatically spotlighted the handful of storm‑time hours that dominate drag errors, giving the model storm “situational awareness” without deep stacks .
Metric‑aligned training: Loss is a time‑weighted MSE whose exponential kernel is identical to the OD‑RMSE leaderboard weights, so 95 % of the gradient signal lives in the first ≈19 h of the horizon . AdamW, gradient clipping to 1.0, AMP, and early stopping land a converged model in ~4 h on a single RTX 3090 Ti; inference clocks in at a few milliseconds.
This end‑to‑end pipeline hits real‑time throughput, outperforms the transformer baseline, and ranks 4th on the Final leaderboard — all while running light enough for flight‑dynamics servers or even on‑board processors.
Results:
MIT STORM-AI Ranking: 4th / 18 validated participants
Phase 1.1 Public (Medium) | Phase 1.1 Public (Hard) | Public Score | Private Score | Model Score | Normalized Model Score | Report Score (Q) |
0.6784 | 0.5398 | 0.5675 | 0.1156 | 0.4998 | 0.7094 | 0.807 |
The model outperformed the official transformer baseline while keeping the parameter count and run‑time budget tiny.
Key Innovations
Metric‑aligned loss – maximizes leaderboard skill instead of generic RMSE.
Per‑sequence density normalization – each file is scaled by its own background density to generalize across altitudes.
Explicit multi‑scale lags (1–12 h) + moving averages – exposes short‑term dynamics without increasing network depth.
Additive attention – automatically up‑weights the handful of storm‑time hours that dominate density variability.
Why It Matters
Operational speed – forecasts arrive fast enough for on‑ground collision‑risk screening or even on‑board drag compensation.
Storm‑time fidelity – exponential weighting + attention means accuracy where operators care most (first 24 h).
Reproducibility – the entire pipeline (data loaders, scalers, PyTorch model) is open‑sourced and deterministic.
What’s Next
I’m would explore three extensions:
Hybrid residuals – add the learned correction on top of NRLMSIS 2.1 to blend physics priors with data‑driven agility.
Uncertainty quantification – last‑layer Laplace + Monte‑Carlo dropout for calibrated confidence intervals.
Continual learning hooks – lightweight fine‑tunes so the model stays sharp through Solar Cycle 25.
Comments