Google DeepMind and Google Research jointly launched WeatherNext 2, a next-generation AI-powered weather forecasting model, this Monday. Hailed as the most advanced and efficient AI weather forecasting system currently available, the model boasts an eight-fold increase in processing speed compared to its predecessor, generating forecasts with resolutions down to the hour. It surpasses traditional methods across 99.9% of meteorological variables (including temperature, wind speed, and humidity) within a 0-15 day forecast period.
The core breakthrough of WeatherNext 2 lies in its innovative multi-scenario forecasting capabilities. Starting from a single state, the model can simulate hundreds of possible weather paths per minute on a single TPU, whereas traditional physics supercomputing models would take hours to complete a similar analysis. This technology has already been applied to experimental cyclone forecasting, assisting meteorological agencies in risk analysis. Its performance leap is attributed to a new AI architecture called Functional Generative Network (FGN), which injects "noise" to ensure the physical plausibility and correlation of factors in the forecast, excelling particularly at handling "marginal" (single-factor) and "joint" (complex system) forecasts in meteorology.
This release marks the first official open application of this technology. WeatherNext 2 data is already integrated with Earth Engine and BigQuery, and is available for early access via Google Cloud's Vertex AI. Google has also integrated it into Search, Gemini, Pixel Weather, and the Weather API on the Google Maps Platform, and will expand to Google Maps weather services in the coming weeks. The company stated that this technology will enable users across various sectors to make more accurate decisions, and will continue to explore expanding data integration and access.
