Google DeepMind published research in Nature detailing WeatherNext, an artificial intelligence model that extends tropical cyclone forecasting accuracy by a full day.
The system predicts storm tracks, intensity, and wind structures simultaneously. According to Google DeepMind, three-day forecasts generated by the model match the accuracy of two-day forecasts from previous systems. The development team worked alongside engineers at Google Research, the National Hurricane Center, the UK Met Office, and the Cooperative Institute for Research in the Atmosphere.
Model Training and Ensembles
Researchers co-trained the system on nearly 20 terabytes of global atmospheric data alongside the IBTrACS database, which contains records for nearly 5,000 historical storms. WeatherNext uses Functional Generative Networks to generate probabilistic ensembles. A single 15-day forecast runs in less than a minute on a TPU.
During the 2025 hurricane season, the system produced 50 predictions at a time and helped forecast Hurricane Melissa’s rapid intensification prior to its landfall in Jamaica. The team scaled the ensemble capacity to 1,000 scenario predictions per storm for the current season to capture rare extreme weather events.
Resolution and Code Availability
Traditional intensity forecasting relies on fine spatial resolution, but WeatherNext Cyclones requires data inputs at a resolution of 28 by 28 kilometers, which is 100 times coarser than legacy physics models. A compact version, WeatherNext 2-mini, processes data at a 111 by 111 kilometer resolution on a single processor in a public Colab notebook. Google DeepMind did not explain the exact mechanisms behind how these models achieve high accuracy at such coarse resolutions.
The company is open sourcing the code and model weights for both WeatherNext 2 and WeatherNext Cyclones. Updated forecasts, including temperature, precipitation, and wind speed layers, are now accessible through Google Earth AI's Weather Lab interface.
