Details
- Google DeepMind, in collaboration with Google Research, has introduced WeatherNext 3, an AI-based global weather forecasting model.
- Unlike traditional systems that typically update forecasts every six hours due to compute constraints, WeatherNext 3 ingests real-time satellite data to generate a fresh global forecast every hour.
- The model is trained directly on raw weather station observations, enabling it to better capture localized microclimates, particularly in historically underserved regions.
- WeatherNext 3 improves temperature forecast resolution by a factor of five, moving from a 25 km grid down to 5 km in a single model pass.
- The system delivers a substantial accuracy gain in global precipitation prediction, with up to a 50% reduction in error compared with previous methods.
- WeatherNext 3 will now underpin weather information across Google products, including Google Search, the Gemini app, Google Maps, and Google Maps Platform.
- Developers and researchers can access real-time WeatherNext 3 forecast data via Google BigQuery, Google Earth Engine, and Google Cloud Storage (GCS), expanding its use in scientific and commercial applications.
Impact
By pushing forecast frequency to hourly and boosting spatial resolution to 5 km, WeatherNext 3 illustrates how frontier AI models are starting to encroach on the domain of traditional numerical weather prediction. Its integration across major Google surfaces and cloud tools is likely to accelerate AI-driven weather services, influencing sectors like logistics, agriculture, and renewable energy that depend on localized, timely forecasts.
