DeepMind's WeatherNext 3 trains on live satellite data instead of six-hour-lagged NWP output
Launched September 3, WeatherNext 3 is a Functional Generative Network mesh transformer that ingests live geostationary satellite mosaics rather than the six-hour-lagged numerical weather prediction data WeatherNext 2 trained on, and it updates hourly at 5km resolution for temperature and moisture against its predecessor's 6-hour, 25km cadence, roughly five times sharper. Medium-range precipitation improves 60% against IMERG satellite data, 30% against MRMS radar and 10% against rain gauges at early lead times, with up to 50% better day-ahead precipitation in regions with sparse ground instrumentation. Developers can query it through BigQuery, Cloud Storage, Earth Engine and the Google Maps Platform Weather API without standing up the model.
↳ Follow the thread