DeepMind's Hurricane AI Bought Forecasters a Day
WeatherNext predicted Hurricane Melissa would hit Jamaica as a Category 5—five days out. That extra day of lead time changes everything.
By The space-quantum Desk
In October 2025, a storm brewed over the Caribbean Sea and weather models couldn't agree on where it was going. Traditional forecasting systems split: some placed the storm on a weakening track toward Haiti; others flagged Jamaica. [1]
Hurricane Melissa did exactly that. Flooding and landslides tore across Jamaica. But because forecasters had the warning earlier, communities had more time to move, stage supplies, and make decisions that matter most when a major storm is hours away. [1]
Melissa also marked a first: it was the first time the National Hurricane Centre was able to predict a Category 5 hurricane when the storm was only at a Category 1 stage—a storm that had intensified rapidly, developing into an emergency situation overnight. On average, the model gives forecasters a day more lead time than existing models; its predictions three days out are as accurate as previous models' predictions two days out. [1]
That gap between knowing and not knowing—between a day gained and a day lost—is where preparation lives and panic dies. In disaster management, the question has never really been about data volume; it has always been about how much signal you can extract from the noise before the window closes.
"Even a few hours can make a difference," says Mike Brennan, director of the US National Hurricane Center. Organizing evacuations, staging supplies, and moving resources to respond to a hurricane risk are all time-sensitive tasks—and making the wrong decision can have big consequences. "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," he says. Historically, bringing forecasts forward by a day would take a decade of work, the researchers say. [1]
The challenge WeatherNext had to solve is structurally awkward. Predicting a hurricane's track—the direction it travels—requires global-scale atmospheric data: cold front positions, prevailing winds, broad pressure systems. Predicting intensity, however, demands something else entirely: much smaller-scale data focused specifically on the local atmospheric and ocean conditions. "That's something we just don't get from these global models," says Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and an author on the paper. Earlier AI models managed storm track well; "intensity they could not do well at all," Musgrave says. [1]
The solution DeepMind and Google Research pursued was architectural: train a model to perform well on both general weather prediction and cyclone-specific behavior simultaneously. "We don't have that much cyclone data, but we have a lot of weather data," says 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." [1]
There is something philosophically strange about a model that outperforms its predecessors while working with less detail. The instinct in complex modeling has always been toward finer resolution, more data, higher-powered compute. WeatherNext suggests that instinct may have been partially wrong—or at least incomplete.
What makes the performance stranger still is the resolution at which it operates. WeatherNext uses lower-resolution atmospheric inputs than traditional numerical models require to forecast storm intensity—a fact that stunned specialists when disclosed. "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 says. The researchers tested the model on retrospective data first, and the numbers were good enough that the team doubted their own results. "The results were so good that we were skeptical that we would actually see that in the real-time demonstration," Musgrave says. When operational use began, the performance held. "I think everybody was surprised at just how well it did," she adds. [1]
The mechanism behind the accuracy remains unclear to the people who built it. The model is picking up on something in coarser data that correlates with intensity outcomes—but what exactly that signal is, no one yet knows. "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 says. Rather than treating that opacity as a problem to be embarrassed about, the DeepMind team frames it as a research lead: evidence that atmospheric physics contains patterns the field has not yet formally described. [1]
WeatherNext also produces outputs differently from conventional models. Rather than a single deterministic forecast, it generates a range of scenarios for each developing storm—a design meant to capture any potential "butterfly effect," says Alet, where a small deviation from a trend could lead to much bigger changes down the line. Last year, the AI model created 50 scenarios per storm; now, it generates 1,000. "That's something that, with our computing power, we simply can't do with our existing numerical models," Musgrave says. Forecasters blend these probabilistic outputs with those of other systems to build their final assessments. [1]
A thousand futures, weighted and mapped, offered to a forecaster who must still make a single call—that is the state of the art right now. The model handles the computation; the judgment belongs to people with meteorological training and an understanding of what flooding means to a specific coastline.
Brennan is careful not to overstate what any single model means for the future of forecasting. "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 says. He is equally clear that the human role in translating model outputs into public safety decisions is non-negotiable. "A hurricane is not just a track or an intensity forecast. It requires experts to translate that into what the impacts are going to be—and it's the impacts that kill people." [1]
Google DeepMind has announced it will open-source the WeatherNext models used during hurricane season, making them available to the broader research community. The move is partly practical—more eyes on the architecture may help explain how coarse-resolution inputs are generating such precise intensity signals. Alet frames it as something more expansive. "I think AI is giving us new tools to poke into the laws of the universe," he says. [1]
What WeatherNext represents is not a replacement for meteorological expertise. It is a compression of the window between ignorance and knowledge—one day thinner, one day earlier. When a storm like Melissa exists only in model output and has not yet touched land, that day is exactly what evacuation logistics, supply pre-positioning, and emergency communication systems run on. For the communities waiting on forecast confidence before they move, a day is not a rounding error. It is the margin. [1]