In October 2025, forecasters diverged on the trajectory of a storm brewing over the Caribbean Sea: would it stay weak and hit Haiti, or intensify and strike Jamaica? According to WIRED, Google DeepMind and Google Research's WeatherNext artificial intelligence model answered with 80 percent confidence, five days before landfall, that it would hit Jamaica as a Category 5 hurricane. The storm, Hurricane Melissa, caused flooding and landslides across Jamaica. The earlier AI-assisted warning helped communities prepare, WIRED reported.
One extra day of lead time
WIRED reported that the WeatherNext model, described in a paper published in Nature, provides forecasters with one day more lead time on average than existing models; its three-day predictions are as accurate as previous models' two-day predictions. The researchers said that, historically, achieving that extra day would take a decade of work.
"Even a few hours can make a difference," Mike Brennan, director of the US National Hurricane Center, told WIRED. "Time is really golden when it comes to those types of decisions, so the ability to push forecast accuracy out as much as a day beyond what we've previously been able to do is really valuable."
Brennan's point reflects the operational realities of hurricane response: organizing evacuations, staging supplies, and moving resources are all time-sensitive tasks, and the wrong decision can have major consequences.
Training on weather data to predict rare cyclones
Machine learning models require ample training data, but extreme events such as cyclones are naturally rare. Ferran Alet, a research scientist at Google DeepMind and a lead author of the paper, explained the approach to WIRED. "We don't have that much cyclone data, but we have a lot of weather data. So what we did was train a model to be both good at weather as well as cyclones."
Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and an author on the paper, told WIRED that hurricanes are particularly difficult to predict because they operate at multiple spatial scales. Predicting a storm's track requires global-scale data, such as the location of cold fronts and prevailing winds; predicting intensity requires much smaller-scale data on local atmospheric and ocean conditions. Earlier AI models handled track well, but, as Musgrave said, "intensity they could not do well at all."
| Prediction target | Data scale required | AI capability before WeatherNext |
|---|---|---|
| Storm track | Global weather data | Good |
| Storm intensity | Local atmospheric/ocean data | Poor |
From retrospective tests to live operations
Before WeatherNext entered live forecasts, researchers tested it on retrospective hurricane data. "The results were so good that we were skeptical that we would actually see that in the real-time demonstration," Musgrave told WIRED. But when forecasters adopted the model into their operations, the performance held. Hurricane Melissa marked the first time the National Hurricane Center was able to predict a Category 5 hurricane while the storm was still at Category 1 strength. "I think everybody was surprised at just how well it did," Musgrave said.
An unexplained accuracy gain
The DeepMind researchers themselves do not fully understand how the model produces such accurate predictions, given that it uses much lower-resolution atmospheric data than traditional models require for forecasting storm intensity. When they told the research community that the model used relatively coarse resolution, Musgrave said, "they were shocked" — a striking outcome for a model that defies conventional assumptions about data resolution requirements.
Sources:
