AI Predicts Cyclones One Day Sooner

Google Deepmind has unveiled WeatherNext Cyclones (WN-C), an advanced AI system designed for tropical cyclone forecasting. This innovative model provides predictions that extend approximately one day beyond the capabilities of existing operational models, marking a significant leap in meteorological forecasting. WN-C was developed in collaboration with the National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere, and the UK Met Office.

Historically, cyclone forecasting has faced a compromise between track and intensity prediction accuracy. Global models excel at track prediction but lack detail on storm intensity, while regional models offer greater intensity precision but sacrifice track accuracy. WN-C resolves this tradeoff by effectively handling both aspects within a single, unified system. This breakthrough was detailed in a recent publication in Nature.

The AI achieves enhanced forecasting accuracy by employing Functional Generative Networks (FGN), a method that allows for a single pass through the neural network, resulting in an eight-fold speed increase compared to diffusion-based methods. WN-C processes data from a relatively coarse grid, yet delivers highly accurate intensity forecasts. This suggests that the coarser weather data contains more information about storm strength than previously understood.

Deepmind has made the code and model weights for WeatherNext 2 and WN-C publicly available on GitHub, enabling broader access and research. The AI is intended to support, not replace, human forecasters. This development follows Deepmind's earlier introduction of WeatherNext 2, its underlying weather AI, and GenCast, a probabilistic weather model that previously surpassed ECMWF's ensemble predictions.