Google's WeatherNext 3 ditches physics simulations and learns weather directly from live satellite data
decoder ·
Google and DeepMind have introduced WeatherNext 3, an AI weather model that bypasses traditional physics-based simulations by learning directly from real-time satellite data. It generates fresh forecasts hourly on a five-kilometer grid, which is roughly five times sharper than its predecessor WeatherNext 2's 25-kilometer, six-hour intervals. This finer detail helps capture local terrain and weather events like narrow rain bands, with significant improvements in regions such as Latin America, Africa, and the Asia-Pacific that were previously underserved. For precipitation predictions, the model shows up to 50 percent more accurate short-range forecasts by training on NASA's IMERG and Google's own radar-based data. It also produces outputs relevant to renewable energy, including wind speeds at turbine height and solar irradiance estimates. The forecasts are now integrated into Google Search, Maps, Gemini, BigQuery, and Earth Engine. Google notes that the atmosphere remains unpredictable and directs users to national weather services for official warnings.
decoder