Google DeepMind and Google Research have announced "WeatherNext 3," a new AI-powered weather forecasting model.
WeatherNext 3 employs a mechanism that directly incorporates the latest observation data obtained from geostationary weather satellites worldwide.
According to the developers, while conventional models experienced a delay of approximately six hours in numerical weather prediction data, WeatherNext 3 can generate new forecasts every hour.
Spatial resolution has also been improved.
While the prediction range in WeatherNext 2 was approximately 25km intervals, WeatherNext 3 is capable of predicting at intervals as small as 5km. This reportedly allows for a more detailed representation of fine temperature differences along the terrain, as well as the shape and extent of rain clouds.
Additionally, NASA's satellite precipitation data and other sources have been used for training.
WeatherNext 3 has already begun being integrated into Google Search, Gemini, and Google Maps.
Researchers and developers can access the data via BigQuery and Google Cloud Storage. Furthermore, an experimental service called "Weather Lab" has been provided, allowing users to verify forecasts on a map.
Source: GoogleがAI気象モデル「WeatherNext 3」発表、衛星データを直接使ってスパコンの待ち時間を回避し1時間ごとに予報更新 (GIGAZINE, 2026-09-04)