AI weather models have spent three years catching up with physical forecasting, but two issues have remained unsolved: coarse resolution and inability to capture local topography; and initialization that lags by about 6 hours behind numerical weather prediction (NWP) analysis. WeatherNext 3, jointly released by Google DeepMind and Google Research, addresses these two points directly.

image.png

Train directly on meteorological station measurements, no longer just look at reanalysis grids

WeatherNext 3 uses real-time global geostationary satellite mosaics as the direct input of the model, which is reinitialized every hour, and outputs a maximum resolution of 0.05° (about 5 kilometers), and it is trained using original meteorological station observation data rather than relying only on reanalysis grids. According to Google AI, Brightband's independent real-time evaluation has rated it as the most accurate global weather model to date.

Architecturally, WeatherNext 3 is a function generation network (FGN) grid Transformer, and it belongs to the same probabilistic model family as WeatherNext 2, but it is extended to multi-resolution output. The inputs include real-time geostationary satellite mosaics and ECMWF HRES analysis. The training data includes ERA5/HRES-fc0, NASA's IMERG, site observations, and satellite mosaics.

Most AI forecast models learn from NWP reanalysis data, which can smooth out local differences caused by coastlines, valleys, and mountains. WeatherNext 3, however, trains directly on original site measurement data using a dedicated observation head, calibrating its 0.05° temperature and dew point outputs to the actual instrument readings.

Hourly updates, significant improvements in precipitation and renewable energy variables

A single forward inference can produce three levels of results: 2-meter temperature and dew point at 0.05° (about 5 km) for site-trained models, 10-meter and 100-meter surface wind, pressure, sea surface temperature, cloud cover, solar radiation, and 1-hour precipitation at 0.1° (about 10 km), and a 0.25° (about 25 km) atmospheric field covering 13 pressure layers. In comparison, WeatherNext 2 outputs a 0.25° field every 6 hours, which is the reason for the claim of "a five-fold clarity improvement."

Refresh frequency is another major change: the model is initialized 24 times a day, with four weather cycles at 00, 06, 12, and 18 UTC able to forecast up to 15 days (360 hours) with 64 ensemble members, while the remaining hourly runs cover 48 hours. For rapidly developing convective weather, the hourly refresh based on current satellite observations makes a substantial difference compared to the 6-hour cycle of lagging analysis.

Precipitation has always been a weakness of global models, with output often blurry and unable to capture storm boundaries. WeatherNext 3 is trained using three data sources: ECMWF reanalysis, NASA IMERG satellite inversion, and Google's own satellite radar precipitation reanalysis. Google claims that, in short-term forecasts, its CRPS improves by up to 60% compared to the baseline (compared to IMERG), 30% (compared to MRMS), and 10% (compared to rain gauges); studies also show that when evaluated against IMERG, the Brier score and CRPS are reduced by up to 50% compared to the NWP baseline.

For renewable energy, the model outputs 100-meter wind speed at the height of turbine hubs, complete low/middle/high cloud distribution, and two components of solar irradiance—SSRD and FDIR—which are exactly the combination required by power grid operators to predict wind and solar output.

Currently, forecast data can be requested through BigQuery, Earth Engine, and Cloud Storage, but the weights of WeatherNext 3 are not open-sourced, and custom inference still runs WeatherNext 2 upon request.