DeepMind's WeatherNext AI Model Enhances Hurricane Forecasting

By Priya Anand

DeepMind has open-sourced its WeatherNext AI model, marking a significant advancement in hurricane forecasting. The model allows forecasters to predict hurricanes with up to an extra day of lead time, using lower-resolution weather data.

WeatherNext demonstrated its capabilities during Hurricane Melissa, where it enabled the U.S. National Hurricane Center to predict the storm's rapid intensification to a Category 5 hurricane and its landfall in Jamaica five days in advance. The forecast was refined to nearly 100% confidence three days before the event.

In tests conducted in 2023 and 2024, WeatherNext's five-day track forecasts were, on average, 140 kilometers closer to actual storm locations compared to the European Centre for Medium-Range Weather Forecasts' ENS ensemble. This improvement equates to a 1.5-day enhancement in forecast accuracy.

Beyond track forecasting, WeatherNext also outperformed NOAA's HAFS model in predicting storm intensity, delivering lower average intensity errors.

The open-source release of WeatherNext includes model code and weights, supporting further research and application by agencies and the broader scientific community. The model can provide forecasts up to 15 days in advance, matching or exceeding traditional benchmarks in various metrics.