Google DeepMind and Google Research say their WeatherNext model can predict tropical cyclones with enough accuracy to give forecasters about one extra day of useful lead time. In hurricane planning, that can change when officials stage supplies, move resources, and decide whether to evacuate people.
The model was tested on retrospective data and then used in live forecasting, according to research published in Nature and reported by WIRED. In one example, WeatherNext predicted that Hurricane Melissa would strike Jamaica as a Category 5 storm five days before landfall, when other models still differed on the storm’s path and intensity.
Cyclones are difficult for AI because rare extreme events provide limited training examples, and storm track and intensity depend on different scales of weather data. The researchers trained WeatherNext to handle both broader weather patterns and cyclone-specific behavior.
The result is not a replacement for human forecasters. Even researchers quoted in the report say they do not fully understand why the lower-resolution model performs so well. Its value is as another forecasting signal that can extend the window for human decisions.