Google DeepMind launches WeatherNext 3 AI weather model
Google DeepMind introduced WeatherNext 3 on Thursday, a weather forecasting model that generates predictions every hour by drawing on live satellite data rather than waiting for government-produced datasets that update every six hours.
The model forecasts wind speeds at 100 meters — roughly the height of a wind turbine — along with cloud cover and solar radiation levels, data the company says will help grid operators and renewable energy developers estimate power output from wind and solar assets. According to DeepMind, the hourly cadence reduces the typical forecast's data lag to three or four hours, down from roughly seven.
Most AI weather models rely on a dataset from the European Centre for Medium-Range Weather Forecasts, one that requires around five hours to compile and is only refreshed every six hours. WeatherNext 3 draws on already-assembled ECMWF output and layers in real-time geostationary satellite imagery to achieve its hourly update cadence.
"It gets much more accurate by not waiting for the next analysis date and using the most recent information," DeepMind senior research scientist Ilan Price told Bloomberg .
WeatherNext 3 also raises the resolution of its forecasts. Key surface variables like temperature and moisture are visualized at a 5-kilometer grid, compared with the 25-kilometer grid and six-hour increments used by its predecessor, WeatherNext 2, the company said. Atmospheric variables including wind speed are forecast at 25-kilometer resolution. The model also trains on sparse weather station observation data, allowing it to target forecasts to specific stations rather than averaged grid cells.
On precipitation, the company said the new model scores up to 60% better on a standard probabilistic accuracy measure compared with WeatherNext 2. The model trains on NASA's satellite-based precipitation dataset as well as a Google-produced precipitation reanalysis.
According to DeepMind, WeatherNext 3 came out on top in ongoing head-to-head testing conducted through Brightband's Operational WeatherBench comparison platform. The platform evaluates models across variables such as temperature, wind speed, and humidity, according to TechCrunch . The model's forecasts have not been independently verified beyond those evaluations, according to Bloomberg.
The model is now powering weather results in Google Search, Gemini, Google Maps, and Google Maps Platform's Weather API, the company said. Forecast data is also available to developers and researchers through BigQuery, Google Earth Engine, and Google Cloud Storage.
WeatherNext 3 has 2.4 times more parameters than WeatherNext 2, according to TechCrunch. DeepMind said the model also natively predicts cyclone tracks alongside its gridded atmospheric outputs.
