WeatherNext 3: hourly satellite AI weather, shipped into Search
Google DeepMind and Google Research launched WeatherNext 3 on Sep 3 — flagship global weather AI that refreshes hourly from live geostationary satellite mosaics, not just lagged NWP reanalysis. It’s wiring into Search, Gemini, Maps, Maps Platform Weather API, and Earth Engine; grids also land in BigQuery / Cloud Storage.
- Resolution vs WeatherNext 2: Surface temp/moisture at 5 km (was 25 km / 6-hour steps); other surface ~10 km; atmospheric (e.g. wind) 25 km — Google’s “~5× sharper” claim.
- Architecture: Functional Generative Network (FGN) mesh transformer; gridded fields, cyclone tracks, native station points. Dev docs: 64-member ensemble; 15-day horizon on 6-hourly cycles, 48 hours on interim hourly inits.
- Precip pitch: CRPS gains vs IMERG up to 60%, MRMS 30%, gauges 10% at early leads; product copy says up to 50% better day-ahead precip where forecasts were weakest.
- Clean-energy vars: 100 m wind, high-res cloud cover, solar radiation — pitched at grid operators and renewables.
- Brightband: Google calls it the most accurate global model “to date” per Brightband live evals — the blog points to the leaderboard, it doesn’t paste the scores.
What it means
Useful: hourly satellite refresh + finer grids + Cloud access is a real ops upgrade over WN2’s six-hour lag. Hype filter: “most advanced” rides a third-party board Google doesn’t reproduce inline, and the post still defers warnings to local mets. Verge notes AI still sits alongside physics models. For builders — query BigQuery/Earth Engine; for consumers — sharper Maps/Search precip, not a replacement for NHC or your national service.