The Google DeepMind team published a blog post yesterday (August 7), announcing the launch of the WeatherNext Cyclones weather forecasting model, designed primarily for cyclones. The team said it delivers industry-leading accuracy in predicting their tracks, intensity, and wind-field structure.
According to the blog post cited by IT Home, the model was jointly researched and developed by Google DeepMind, Google Research, the U.S. National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and multiple meteorological agencies worldwide. The model focuses on tropical cyclones and provides highly accurate predictions of the tracks, intensity, and wind-field structure of typhoons, hurricanes, and cyclones.
Google said that, measured by average performance, the model advances forecasting capabilities by 24 hours. Existing models can typically make predictions only 2 days in advance, while this model can forecast 3 days ahead while maintaining accuracy. In its blog post, Google described the scale of the forecast as “about 10 years of meteorological progress.”
For training, the model learned end-to-end from nearly 20 TB of global atmospheric data, as well as the IBTrACS (International Best Track Archive for Climate Stewardship) database covering nearly 5,000 historical storms.
In terms of generation capabilities, with the help of Functional Generative Networks, the system can generate a single 15-day forecast on a TPU in less than 1 minute. Last year, the system generated only 50 forecasts at a time; this has now been expanded to 1000.
In practical applications, WeatherNext predicted during the 2025 hurricane season that Hurricane Melissa would intensify and make landfall in Jamaica. The NHC issued an early warning based on the prediction, giving local emergency teams more time to prepare.
The Google team has also open-sourced WeatherNext 2, WeatherNext Cyclones, and WeatherNext 2-mini, along with the related code and model weights.