DeepMind AI predicts hurricanes a day earlier than existing models

An AI model developed by Google DeepMind can predict hurricanes with a full day more lead time than existing forecasting systems, according to research published in Nature on August 6, 2026. The WeatherNext model delivered three-day forecasts as accurate as previous models’ two-day predictions, giving forecasters earlier warnings for devastating storms. The extra lead time helped communities prepare for Hurricane Melissa, which caused catastrophic flooding and landslides across Jamaica.

Mike Brennan, director of the US National Hurricane Center, said even a few hours of extra warning can be critical for organizing evacuations and staging supplies. “Time is really golden when it comes to those types of decisions,” he noted, calling the ability to push forecast accuracy out by a day “really valuable.” The researchers say that historically, such a leap would have required a decade of work.

Hurricanes are notoriously difficult to predict because they involve multiple spatial scales. Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a co-author, explained that predicting a storm’s track requires global-scale weather data, while intensity forecasts need much finer local data on atmospheric and ocean conditions. Previous AI models struggled with intensity, but WeatherNext succeeded. During Hurricane Melissa, the National Hurricane Center was able to predict a Category 5 storm when it was still only a Category 1—a first.

Ferran Alet, a research scientist at Google DeepMind and a lead author, said the team trained the model on vast amounts of general weather data to compensate for the scarcity of cyclone data. The model uses relatively coarse-resolution atmospheric inputs, yet it captures signals about storm intensity that surprised scientists. “When we told the community that our model was only using relatively coarse resolution, they were shocked,” Alet said. In 2026, the model generates 1,000 potential scenarios per storm, up from 50 in 2025, helping forecasters account for the butterfly effect.

Before live deployment, the model was tested on retrospective data with results so strong that researchers were skeptical they would hold in real time. Musgrave said the performance did hold true, surprising everyone. However, Brennan cautioned that the AI model is one tool among many and no single model is guaranteed to be best for every future storm. The human element remains essential to translate forecasts into impact warnings that save lives.

Google DeepMind announced it is open-sourcing the WeatherNext models used during hurricane season, allowing researchers worldwide to use and improve them. Alet expressed hope that this will lead to new scientific discoveries about cyclone behavior, saying, “I think AI is giving us new tools to poke into the laws of the universe.”