Google DeepMind's AI Gave Jamaica an Extra Day to Prepare for a Category 5 Hurricane
A new storm-forecasting model predicted Hurricane Melissa would strike Jamaica five days out, a full day earlier than older tools could manage. A paper in Nature says the AI works in ways even its own creators cannot fully explain.
Key points
- Google DeepMind's WeatherNext model predicted Hurricane Melissa would hit Jamaica as a Category 5 storm five days before landfall, with 80 percent confidence.
- On average, WeatherNext gives forecasters one full extra day of accurate warning compared with existing models, according to a paper published in Nature on Thursday.
- The model generated 1,000 possible storm scenarios at once, a volume traditional computing cannot match.
- DeepMind is releasing the WeatherNext models as open-source software so outside researchers can study and improve them.
- Even the model's own developers cannot fully explain how it achieves its accuracy using lower-quality input data than traditional forecasting requires.
In October 2025, a storm built quietly over the Caribbean. Traditional weather models disagreed: would it stay weak and drift toward Haiti, or strengthen and turn toward Jamaica? WeatherNext, an AI forecasting system built by Google's DeepMind and Google Research divisions, sided firmly with the second option. Five days before the storm made landfall, it flagged an 80 percent chance of a Category 5 hurricane, the most destructive class, striking Jamaica.
It was right. Hurricane Melissa brought flooding and landslides across the island. But the early warning meant communities had more time to prepare.
What makes this different from regular weather forecasting?
WeatherNext buys forecasters roughly one extra day of reliable warning. That means its three-day forecast is about as accurate as older models' two-day forecasts. Historically, the researchers say, gaining that single day would have taken a decade of incremental scientific work.
Mike Brennan, director of the US National Hurricane Center, put it simply: "Time is really golden when it comes to those types of decisions."
Organising evacuations and moving supplies are time-sensitive jobs. Getting them wrong costs lives. One more day matters enormously.
Why are hurricanes so hard to predict in the first place?
Two things must be predicted at once, and they demand very different kinds of data. Tracking which direction a storm travels requires big-picture global weather data: cold front positions, wind patterns across whole ocean basins. Predicting how strong the storm will get requires fine-grained local data about ocean temperature and the atmosphere directly above the storm. Earlier AI models handled track reasonably well. Intensity, says Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a co-author on the paper, "they could not do well at all."
DeepMind's team solved part of this by training WeatherNext on huge volumes of general weather data alongside the far smaller set of historical cyclone records. The model learned broad atmospheric patterns first, then applied that knowledge to storms.
The strangest twist: the system uses lower-resolution input data than traditional models consider necessary for intensity forecasting, yet it still outperforms them. When the team told other scientists this, "they were shocked," says Ferran Alet, a research scientist at Google DeepMind and a lead author on the paper. Something in the coarser data carries more useful signal than anyone previously suspected. Nobody yet knows what.
"It's a black box at the end of the day," Alet acknowledges.
What happens next?
DeepMind is open-sourcing the WeatherNext models so that independent researchers can probe them and, potentially, figure out what the AI spotted that human physicists missed.
The model now generates 1,000 possible storm scenarios for each developing hurricane, up from 50 last year. Traditional numerical models, which run physics equations step by step, cannot get anywhere near that number with available computing power.
Brennan is clear-eyed about the limits. No single model is guaranteed to perform best in every future season. Human experts still translate raw forecasts into impact assessments, the part that ultimately determines how communities respond. "It's the impacts that kill people," he says.
WeatherNext is a sharper instrument in a large toolkit. It is not a replacement for the people who wield it.
