Google DeepMind's WeatherNext predicts cyclone track and intensity simultaneously

Google DeepMind's WeatherNext Cyclones (WN-C) is an AI system for predicting tropical cyclones that can forecast about a day further ahead than leading operational models. Developed with the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere, and the UK Met Office, it has been running live on Google's Weather Lab since June 2025. For a five-day forecast, the average position error is 230 kilometers, compared to 370 for ECMWF's ensemble (ENS) and 335 for DeepMind's previous GenCast model. For three-day intensity predictions, WN-C is 3.75 knots more accurate than NOAA's HAFS. Despite using a grid where each point covers about 28 square kilometers, roughly 100 times coarser than specialized regional models, the system performs better, leading researchers to conclude that high resolution is not strictly necessary for state-of-the-art intensity forecasts. The model uses Functional Generative Networks (FGN) instead of diffusion, making it eight times faster. It was trained on nearly 20 terabytes of global atmospheric data and a curated database of about 5,000 historical cyclones. The system can generate 1,000 parallel forecast scenarios per storm, helping to capture rare extreme events. While WN-C improves track forecasts by 28% when combined with consensus models, it only adds about 6% improvement for intensity, indicating that classical models still contribute significantly. DeepMind has made the code and weights for WeatherNext 2 and WN-C freely available on GitHub, and a mini version can run in a free Colab notebook. However, the authors emphasize that national weather services remain authoritative for official warnings.