DeepMind’s WeatherNext gives forecasters a full extra day of cyclone warning

8:00 AM PT · August 11, 2026

Google DeepMind detailed WeatherNext, a machine learning model for tropical cyclone forecasting, in a paper published in Nature. The model is built to bridge two problems that forecasting systems have traditionally handled separately: modeling broad global weather patterns and modeling the fine scale, localized dynamics of an individual cyclone, combining both in a single system. DeepMind says the result is more than a full day of additional lead time compared with existing operational forecasts. WeatherNext’s three day forecasts of a storm’s track, intensity, and wind structure are roughly as accurate as what prior models could only achieve two days out, a jump the team compares to about a decade’s worth of typical progress in operational meteorology. Because cyclones are among the costliest and deadliest natural disasters, even modest gains in lead time can meaningfully change evacuation planning and disaster response, giving emergency managers more time to act before landfall. DeepMind has released the model’s code and weights openly on GitHub rather than keeping it proprietary, following a pattern the lab has used for other applied science tools such as its earlier flood forecasting systems, in the hope that national weather agencies and researchers will adopt and build on it directly.