Google DeepMind and Google Research have released WeatherNext 3, an AI weather model that delivers global forecasts at 0.05° (~5 km) resolution and refreshes every hour. The model addresses previous limitations by using live geostationary satellite mosaics for initialization instead of lagged numerical weather prediction analysis, and it trains directly on raw weather station observations.
- WeatherNext 3 is a Functional Generative Network mesh transformer that produces multi-resolution outputs in a single forward pass: 0.05° temperature/dew point, 0.1° surface variables, and 0.25° atmospheric fields.
- It initializes 24 times daily, with synoptic cycles running out to 15 days (360 hours) with 64 ensemble members and hourly runs covering 48 hours.
- Precipitation training uses ECMWF reanalysis, NASA’s IMERG, and Google’s satellite-radar data, achieving up to 60% CRPS improvement against IMERG at early lead times.
- The model outputs 100 m wind speed, cloud distributions, and solar irradiance components for renewable energy forecasting.
Forecast data is available via BigQuery, Earth Engine, and Cloud Storage upon allowlist request, though the weights are not open source. Independent evaluations from Brightband rank it as the most accurate global weather model to date.