The question: when the Sun acts up, the upper atmosphere swells and satellites feel more drag. Which solar and geomagnetic drivers move the air in each specific place — and how sure are we?
Air-density response through 40 storms (2021–2025), read from the accelerometers and GPS navigation residuals of two independent satellite fleets that never shared a model: GRACE-FO (460–510 km) and ESA Swarm (420–460 and 510–550 km) — 24 million Swarm measurements across 48 cells, learned one storm at a time with the answer hidden (leave-one-storm-out).
In a third of the storm sectors we can name the drivers that move the air, with most of that learned weight confirmed by repeated correct prediction — and a deliberately scrambled decoy signal correctly earns nothing there.
Where the learner beats simple references is not random — dawn and midnight sectors win, noon does not — and the same pattern appeared on the second fleet, which the model had never seen. Where the learner is not confident, its abstentions correctly flagged its own worst cells before they happened.
A head-to-head win over the standard density model (MSIS) once that model gets the same local calibration any operator applies. We tested that claim ourselves and it did not survive — so it is not on the map. What is ours is the driver attribution and the where-the-reference-is-blind map.
| gap | what closes it |
|---|---|
| 11 cells measure fine but trip the decoy alarm — something in the data can fool a learner there | resolving the data-quality question in those cells; until then, no confidence claimed |
| 120 cells have no source satellite at all (550–850 km especially) | the drag residual any operating satellite already computes for navigation — not telemetry, not proprietary state |
| severe-storm hours are rare — storm confidence grows in calendar time | the next solar-cycle declining-phase season, already scheduled |
This grid grew out of the original LEO density campaign (GRACE-FO era) — the workshop poster from that campaign is preserved here.
GRACE-FO accelerometer-derived densities (TU Delft TOLEOS) · ESA Swarm A/B/C · NASA OMNI drivers · storm-fold definitions and every threshold pre-registered.
Every claim on this page comes from a pre-registered analysis: the hypothesis and its pass/fail rule were committed to a ledger before the data was touched, results were reproduced bit-for-bit, decoy signals ran as tripwires, and refused hypotheses stay on record. The complete gap register — data sources, refused hypotheses, and open staked predictions — is published alongside this atlas (how to read it).
Nervous Machine