DeepMind's WeatherNext model improves cyclone forecasting accuracy
WeatherNext gives roughly an extra day of predictive accuracy on cyclone track, intensity and wind structure versus prior forecasting models, per a Nature paper.
- Models & capabilities
- Notable
Google DeepMind and Google Research published a paper in Nature, developed with the US National Hurricane Center, the UK Met Office and the Cooperative Institute for Research in the Atmosphere, describing WeatherNext, a model that predicts tropical cyclone track, intensity and wind structure from a single unified system rather than the separate models forecasters had previously needed for large-scale steering patterns and localised storm dynamics.
The reported gain was roughly a day of additional warning time: three-day WeatherNext forecasts matched the accuracy that the best prior models achieved only two days out, across track, intensity and wind-structure predictions alike. The model ran at a resolution of about 28 by 28 kilometres — roughly a hundred times coarser than some traditional physics-based forecasting systems — while still outperforming them, and was trained on around 20 terabytes of atmospheric data together with records from close to 5,000 historical storms. DeepMind said the model had generated 1,000-member ensemble forecasts during the 2025 hurricane season and had correctly anticipated Hurricane Melissa’s rapid intensification and landfall in Jamaica.
DeepMind released the underlying code and weights on GitHub under an Apache 2.0 licence, in three variants — WeatherNext Cyclones, WeatherNext 2 and a smaller WeatherNext 2-mini light enough to run on a single TPU inside a free Google Colab notebook — making the tools available to national weather services and researchers without DeepMind’s own compute. The release extended a run of DeepMind weather and climate models, following earlier systems such as GraphCast, into a domain — cyclone forecasting — where lead time translates directly into evacuation decisions.