Google opens advanced AI cyclone forecasting model

Google has unveiled an advanced artificial intelligence cyclone forecasting system, offering researchers worldwide access to its WeatherNext 2 technology and model weights. This breakthrough, detailed in a Nature publication, delivers up to a 24-hour advance warning over current operational forecasting methods for storm tracks, intensity and wind radii.

The WeatherNext Cyclones component specifically outperformed leading models in predicting tropical cyclones from 2023-2025, with each prediction providing roughly a day more lead time than conventional models issued a day later. The system's success stems from merging global atmospheric forecasting with specialized knowledge of tropical cyclones, utilizing nearly 20 terabytes of atmospheric data and historical storm records.

Despite operating at a 28km by 28km resolution, WeatherNext Cyclones generated over 1,000 forecast scenarios, enabling forecasters to assess probabilities and plan accordingly. Google's technology, powered by Functional Generative Networks, can produce 15-day forecasts in under a minute and has already been integrated into various Google services, including Search, Gemini, Pixel Weather and Google Maps.

While official warnings still fall under national meteorological authorities' jurisdictions, the open-source release aims to democratize advanced ensemble forecasting for institutions with limited access to costly supercomputing resources.