Google DeepMind and Google Research have launched WeatherNext 3, an AI weather forecasting model that uses real-time satellite observations to produce global forecasts every hour at resolutions of up to five kilometres.

The model combines live geostationary satellite imagery with historical weather data and observations from weather stations. Compared with WeatherNext 2, which produced forecasts on a 25-kilometre grid every six hours, the new system can provide surface variables such as temperature and moisture at five-kilometre resolution, while other atmospheric variables are generated at different resolutions. Google says independent live evaluations by Brightband place WeatherNext 3 among the strongest-performing global weather models.

WeatherNext 3 also includes variables designed for renewable energy forecasting. It can estimate wind speeds at around turbine height, as well as cloud cover and solar radiation, allowing wind and solar operators to better estimate the expected electricity generation and match supply with demand. Google also reports improvements in precipitation forecasting, although the performance figures come from evaluations cited by the company.

The model is being integrated into Google Search, Gemini, Maps, Google Maps Platform, Earth Engine, and Cloud. Forecast data will also be available to researchers, developers, and businesses through BigQuery, Earth Engine, and Google Cloud Storage.

Why does it matter?

More frequent and higher-resolution forecasts can improve planning for extreme weather, agriculture, and renewable energy generation. WeatherNext 3 also shows how AI and satellite data are being combined to make detailed forecasting available in regions where traditional high-resolution models can be costly to run.