WeatherNext 3 Hits 5 km Resolution With Hourly Satellite Initialization

Google DeepMind and Google Research have released WeatherNext 3, a global weather model that resolves surface conditions to 0.05° (~5 km) and re-initializes every hour by ingesting a live geostationary satellite mosaic — directly addressing two structural weaknesses that have limited AI forecasters: resolution too coarse for local terrain effects, and initialization locked to NWP analysis that arrives with roughly a six-hour lag. Brightband's independent live evaluations rank it the most accurate global weather model currently running. Access is still gated: forecast data is available through BigQuery, Earth Engine, and Cloud Storage via allowlist request; model weights are not open, and on-demand custom inference continues to run WeatherNext 2.

Architecture and Inputs

WeatherNext 3 is a Functional Generative Network (FGN) mesh transformer — the same probabilistic family introduced with WeatherNext 2 — extended to produce multi-resolution output in a single forward pass. The input stack combines a live global geostationary satellite mosaic with ECMWF HRES analysis. Training draws on ERA5/HRES-fc0 reanalysis, NASA's IMERG precipitation retrievals, raw weather station measurements, and satellite mosaics.

The critical departure is the observational training heads. Most AI weather models learn from NWP reanalysis grids, which smooth away local variation produced by coastlines, valleys, and mountain ranges. WeatherNext 3 trains dedicated heads directly against raw station observations, calibrating its 0.05° temperature and dew point outputs to what instruments actually record rather than to a model's representation of the atmosphere.

Resolution Tiers and Forecast Cadence

A single forward pass produces three output tiers. The 0.05° (~5 km) tier covers station-trained 2 m temperature and dew point. The 0.1° (~10 km) tier delivers gridded surface wind at 10 m and 100 m, pressure, sea surface temperature, cloud layers, solar radiation, and 1-hour precipitation accumulation. The 0.25° (~25 km) tier covers atmospheric fields across 13 pressure levels. WeatherNext 2 produced only 0.25° fields in 6-hour increments, making the finest tier roughly 5× sharper spatially and the cadence 6× finer temporally.

The model initializes 24 times per day. The four synoptic cycles — 00, 06, 12, and 18 UTC — extend to a 15-day (360-hour) horizon with a 64-member probabilistic ensemble. Interim hourly runs cover 48 hours. For fast-developing convection, an hourly update grounded in current satellite observations is operationally distinct from a six-hourly cycle anchored to lagged analysis. This is why pipeline architecture choices drive practical AI gains that model scale alone cannot replicate: the data-ingestion design here does work that a larger version of the old architecture could not.

Precipitation and Clean Energy Variables

Precipitation is where global models have historically degraded most severely, producing diffuse fields that miss storm boundaries. WeatherNext 3 trains against three precipitation sources simultaneously: ECMWF reanalysis, NASA's IMERG, and Google's own satellite-radar precipitation reanalysis. The gains above are Google's reported figures; the multi-source training approach produces the range across evaluation datasets rather than overfitting to a single reference.

The renewable-energy variable set signals that this release targets operational buyers. Outputs include 100 m wind speed at approximate turbine hub height, full low, medium, and high cloud distributions, and both solar irradiance components — surface solar radiation downwards (SSRD) and direct solar radiation (FDIR). Grid operators forecasting wind and solar output against demand curves need exactly this combination.

Whether the weights eventually open, and whether the allowlist access path expands to broader API availability, will determine how quickly third-party operators can build on the precipitation and clean-energy outputs that define this model's commercial case.