Google Deepmind has introduced WeatherNext Cyclones (WN-C), an AI system that forecasts tropical cyclone tracks and intensity simultaneously. The model outperforms specialized models in accuracy, offering better predictions for storm paths and strength. WN-C, developed in collaboration with the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), and the UK Met Office, has been running live on Google's Weather Lab since June 2025. During Hurricane Melissa in 2025, the model enabled the NHC to predict the storm's rapid intensification, a critical factor in disaster preparedness.
WN-C addresses a long-standing challenge in cyclone forecasting by integrating track and intensity predictions into a single system. For a five-day forecast, the model's estimated storm center position is off by an average of 230 kilometers, compared to 370 kilometers for the ECMWF ensemble system (ENS) and 335 kilometers for Deepmind's previous model, GenCast. On three-day intensity forecasts, WN-C is 3.75 knots more accurate than NOAA's Hurricane Analysis and Forecast System (HAFS). The model advances the global atmospheric state in 6-hour steps and derives cyclone tracks directly, providing probabilistic forecasts of wind speeds at various thresholds.
The model's success stems from its use of Functional Generative Networks (FGN), which replace the diffusion method used by GenCast. FGN allows for faster predictions with a single pass through the network, making the model eight times quicker. Training on nearly 20 terabytes of global atmospheric data and a curated database of 5,000 historical cyclones enables the model to learn from both sources simultaneously. The authors note that the coarse data used by WN-C contains more information about storm strength than previously thought, though the exact mechanisms remain an open research question.
Source: thedecoder
