In October 2025, a storm system forming over the Caribbean could have followed two very different paths. Traditional weather models disagreed. Google DeepMind’s AI model WeatherNext did not hesitate: five days before landfall, it predicted with 80 percent confidence that the storm would strike Jamaica as a Category 5 hurricane. It was right. Hurricane Melissa caused widespread flooding and landslides across the island, but communities received earlier warnings than would otherwise have been possible. That head start came from a system that, according to research published in Nature, can give forecasters an average of one full day more lead time than existing models.
One extra day sounds modest. In hurricane forecasting, it is not.
The Gap Between Track and Intensity
For years, AI weather models have shown genuine promise at predicting a storm’s track, meaning the direction it is likely to travel. That part of the problem is relatively tractable: it draws on large-scale atmospheric data, the kind that global models handle well. Cold fronts, prevailing winds, broad pressure systems all leave clear signals in the data.
Intensity is a different problem entirely. Predicting how strong a storm will become requires understanding conditions at a much smaller spatial scale: local ocean temperatures, atmospheric dynamics close to the storm’s core, subtle interactions that global models tend to smooth over. As Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and a co-author of the Nature paper, puts it, intensity is something earlier AI models “could not do well at all.”
This is what makes WeatherNext’s performance notable. The model handles both track and intensity, and it does so using lower-resolution atmospheric data than traditional numerical models typically require for intensity forecasting. When the research team disclosed this to the broader meteorological community, the reaction was surprise. The assumption had been that coarse-resolution inputs simply could not carry enough signal to forecast storm intensity reliably. WeatherNext suggests that assumption was wrong, though the researchers themselves cannot yet explain why. Ferran Alet, a research scientist at Google DeepMind and one of the paper’s lead authors, describes the model plainly: “It’s a black box at the end of the day.” Something in the lower-resolution data is informative in ways that were not previously understood. What exactly that something is remains an open question.
A Thousand Scenarios Instead of One
Part of what makes the model operationally useful is not just its accuracy on a single prediction, but the range of possibilities it generates. Rather than producing one forecast, WeatherNext outputs multiple potential scenarios for a developing storm, capturing what Alet describes as the “butterfly effect”: small early deviations that can compound into dramatically different outcomes. Last year, the model generated 50 scenarios per storm. It now generates 1,000. That volume of ensemble forecasting is computationally out of reach for traditional numerical models.
Hurricane Melissa also marked a specific milestone: it was the first time the National Hurricane Center was able to predict a Category 5 hurricane while the storm was still at Category 1 intensity. That kind of early-stage intensity prediction is precisely what the model was designed to improve.
The performance held up under real-world conditions, which was not guaranteed. Before WeatherNext was used in live forecasts, researchers tested it on historical data. The results were strong enough that the team was skeptical they would replicate in real time. They did. “I think everybody was surprised at just how well it did,” Musgrave said.
Why One Day Changes Everything on the Ground
The significance of an extra day of lead time is not abstract. Mike Brennan, director of the US National Hurricane Center, describes the practical stakes clearly: organizing evacuations, positioning supplies, and moving emergency resources are all time-sensitive operations where a wrong call carries serious consequences. Historically, advancing forecast accuracy by a single day would have required roughly a decade of incremental work in traditional modeling. WeatherNext compressed that timeline.
This is where the broader meaning of the technology becomes visible. AI is not replacing the forecasters who interpret these models or the emergency managers who act on them. Brennan is explicit on this point: a hurricane forecast is not just a track or an intensity number. Translating that data into an understanding of real-world impacts, the kind of judgment that determines how communities prepare and respond, still requires human expertise. The model is a tool in a larger system, and a powerful one, but the human layer remains essential.
Google DeepMind has announced it will open-source the WeatherNext models used during hurricane season, making them available to the broader research community. Alet frames this as an opportunity not just for better forecasts, but for scientific discovery. A model that finds signal in data where physicists did not expect to find it may, once examined more closely, reveal something genuinely new about how cyclones work.
In Short
WeatherNext gives hurricane forecasters an average of one extra day of accurate lead time, a gain that historically would have taken a decade to achieve through conventional methods. It does this by handling both storm track and intensity, using lower-resolution data than experts believed possible. The model generates 1,000 scenarios per storm, capturing uncertainty in ways traditional computing cannot match. What it cannot yet do is explain itself: the mechanism behind its accuracy remains unknown, which is itself a scientific signal worth investigating. The human forecasters who interpret its outputs remain indispensable.
Based on reporting from Wired.