Google DeepMind and Google Research are publishing new WeatherNext research in Nature. WeatherNext is an AI weather model that predicts tropical cyclones' paths, intensity, and wind fields. In testing, it achieved the same accuracy as existing models while giving forecasters about 24 hours more lead time, on average. A forecast made three days out, for example, matched the accuracy of other models' two-day forecasts.
The improvement is mostly in intensity forecasting. AI models have long been good at predicting where a hurricane will go, but not how strong it will ultimately become. For disaster warnings, knowing a storm will make landfall isn't enough — you need to know as early as possible whether it'll hit as a Category 1 or a Category 5.
During the 2025 hurricane season, the U.S. National Hurricane Center has been using WeatherNext as a real-time forecasting reference. Five days before Hurricane Melissa made landfall in Jamaica, WeatherNext assigned an 80% probability that it would arrive as a Category 5. Three days out, that probability approached 100%. It also marked the first time the National Hurricane Center forecast that a hurricane would intensify to Category 5 while it was still only a Category 1.