DeepMind's AI Predicts Hurricanes a Day Earlier Than Current Models

In October 2025, a tropical system formed over the Caribbean Sea with an uncertain future. Traditional weather models disagreed on its path—would it weaken and drift toward Haiti, or intensify and strike Jamaica? WeatherNext, an AI model developed by Google DeepMind and Google Research, predicted the latter. Five days before landfall, the model assigned 80 percent confidence to a Category 5 hurricane strike on Jamaica. Hurricane Melissa materialized exactly as predicted, becoming catastrophic and causing flooding and landslides, but the early AI-assisted warning allowed communities in its path to prepare more effectively.

On Thursday, researchers published findings in Nature showing that WeatherNext achieves unprecedented hurricane prediction accuracy. The model delivers, on average, a full day of additional lead time compared to existing forecasting systems. Put another way, its three-day forecasts are as accurate as traditional models' two-day forecasts. Mike Brennan, director of the US National Hurricane Center, underscored the operational significance: "Even a few hours can make a difference" when organizing evacuations, staging supplies, and deploying emergency resources. "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." The researchers noted that achieving a one-day forecast improvement through conventional model development would historically have required a decade of work.

The challenge WeatherNext overcomes centers on predicting both hurricane track and intensity simultaneously. Forecasting track—the direction a storm travels—requires global-scale atmospheric data: the positions of cold fronts, prevailing wind patterns, and other planetary-scale signals. Predicting intensity, however, demands fine-grained local data capturing atmospheric and ocean conditions near the storm itself. As Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a lead author, explained: "That's something we just don't get from these global models. While earlier AI models have done well at predicting a storm's track, intensity they could not do well at all." The distinction matters profoundly—intensity changes can transform a relatively weak system into a major hurricane overnight, as Hurricane Melissa demonstrated when it developed from a Category 1 to a Category 5 hurricane, a feat the National Hurricane Centre had never before predicted from such an early stage.

Training an AI model on hurricanes is inherently difficult because extreme events are rare. Ferran Alet, a research scientist at Google DeepMind and lead author, explained the solution: "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." The approach proved so effective that researchers themselves were initially skeptical. "The results were so good that we were skeptical that we would actually see that in the real-time demonstration," Musgrave said. Yet when forecasters deployed the model operationally, the performance held. "I think everybody was surprised at just how well it did."

Even more intriguing: the researchers do not fully understand how WeatherNext produces such accurate intensity predictions, given that it operates on much lower-resolution atmospheric data than traditional models require. "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. The model appears to detect patterns in coarse data that predict storm behavior, but the mechanism remains opaque. "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." Rather than outputting a single forecast, WeatherNext generates multiple scenarios to capture potential "butterfly effect" dynamics, where small deviations cascade into large changes. Last year the model produced 50 scenarios per storm; now it generates 1,000—a computational capacity Musgrave noted traditional numerical models cannot match. Brennan cautioned that no single model guarantees success across seasons or storms and that human forecasters remain essential for translating track and intensity predictions into impact assessments. "It's the impacts that kill people," he said. Google DeepMind announced it will open-source the WeatherNext models used during hurricane season, inviting researchers to study and refine them. Alet expressed hope that opening the models to the scientific community could reveal fresh insights into cyclone physics. "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."