Dynamic Beating AI News Flash: Google's DeepMind and Google Research have released WeatherNext 3, which directly integrates global real-time satellite observations to regenerate a global weather forecast every hour. The previous generation, WeatherNext 2, was primarily based on a 25-kilometer grid updated every 6 hours; the new model achieves a grid resolution of about 5 kilometers for surface variables like temperature and humidity.

Rainfall and snowfall are the key areas of improvement in this release. WeatherNext 3 incorporates NASA satellite precipitation and radar data for training. Google stated that in scenarios where users need forecasts a day or more in advance, the accuracy of precipitation forecasts could be improved by up to 50%. Previous AI weather models often struggled to predict localized heavy rainfall accurately, so the focus this time is on enhancing the details of such rapidly changing weather.

WeatherNext 3 has started integrating with Google Search, Gemini, Maps, Google Maps Platform Weather API, and Earth Engine. Developers can also access forecast data through BigQuery, Earth Engine, and Google Cloud Storage.

Traditional physics-based weather models will not disappear immediately. The training of WeatherNext 3 still relies on data from such models, and meteorological agencies will consolidate multiple forecasts when issuing disaster warnings. Google also emphasizes that severe weather and public safety alerts should still be relied upon from local meteorological departments.