Five days out from landfall, an AI model put 80 percent confidence on Hurricane Melissa striking Jamaica as a Category 5. Not Haiti. Not a weakened system. Jamaica, at the very top of the scale.
That forecast came in October 2025, at a moment when conventional models were still divided over whether the Caribbean storm would intensify at all, and it is the strongest proof yet of what Google DeepMind and Google Research have built. The system is called WeatherNext, and the researchers behind it cannot fully explain why it works.
The storm was catastrophic regardless. Flooding, landslides, a wrecked Jamaica. What changed is that forecasters pushed warnings out sooner to the communities in its path, which is the entire point of a forecast.
A day of lead time used to cost a decade
Writing Thursday in Nature, the researchers report that WeatherNext forecasts cyclones with an accuracy that beats existing models by roughly a day. What it produces at three days is about as reliable as what older models delivered at two.
Buying that extra day has historically taken around 10 years of work, the researchers said.
“Even a few hours can make a difference,” said Mike Brennan, director of the US National Hurricane Center. Evacuations, staging supplies, moving response resources: all of it runs on a clock, and a wrong call is expensive.
“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 said.
The intensity problem nobody had solved
Track and intensity are two separate forecasting challenges, and AI had managed to solve only one of them.
A storm’s path is governed by planetary-scale weather — cold fronts, prevailing winds, the global picture. Its strength depends on local atmospheric and ocean conditions at a far finer scale, said Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and an author on the paper.
“That’s something we just don’t get from these global models,” Musgrave said. Track was handled well by earlier AI models, but “intensity they could not do well at all.”
Both matter, because intensity is what separates a nuisance storm from a major hurricane. With Melissa, the National Hurricane Centre called a Category 5 for the first time while the system was still a Category 1.
Training on weather to predict cyclones
Machine learning runs on examples. Extreme events, by their nature, are in short supply.
“We don’t have that much cyclone data, but we have a lot of weather data,” said Ferran Alet, a research scientist at Google DeepMind and one of the paper’s lead authors. “So what we did was train a model to be both good at weather as well as cyclones.”
Before the model went anywhere near a live forecast, the team tested it against retrospective data. “The results were so good that we were skeptical that we would actually see that in the real-time demonstration,” Musgrave said. Then forecasters put it to work operationally, and the numbers held up. “I think everybody was surprised at just how well it did,” Musgrave said.
Coarse data, better answers, no explanation
This is the part that rattled people: the atmospheric data WeatherNext works from is far lower in resolution than what traditional models require for intensity forecasting, and it comes out ahead anyway.
“When we told the community that our model was only using relatively coarse resolution, they were shocked, because that means that the lower-resolution inputs capture more signal about what’s going to happen than previously believed,” Alet said.
Something in that coarse data is being picked up by the model. Precisely what, nobody can say. “It’s a black box at the end of the day, but that gives physicists a signal that something is happening that was not previously understood,” Alet said.
From 50 scenarios to 1,000
Rather than a single answer, WeatherNext puts out a spread of possible futures for a developing storm — the way to catch a butterfly effect, Alet said, in which a slight early deviation snowballs into something wholly different days down the line.
A year ago the model was generating 50 scenarios per storm. It now generates 1,000.
“That’s something that, with our computing power, we simply can’t do with our existing numerical models,” Musgrave said.
One tool, not the tool
Brennan is cautious about leaning too hard on any one model, this one included. “There’s no guarantee that one model, because it did well last year or really did well for this particular storm, is necessarily going to be the best model for the next season or the next storm,” he said.
Nor do the humans get taken out of the loop. “A hurricane is not just a track or an intensity forecast,” Brennan said. “It requires experts to translate that into what the impacts are going to be, and it’s the impacts that kill people.”
The WeatherNext models used through hurricane season are being open-sourced by Google DeepMind, letting other researchers run them and build on top of them. Alet is hoping outside eyes will locate the physics hidden inside the black box.
“I’m very excited about scientific discovery,” he said. “I think AI is giving us new tools to poke into the laws of the universe.”