Hurricane forecasting is a race against time. According to researchers writing in Nature, WeatherNext, an artificial intelligence model developed by Google DeepMind and Google Research, may give forecasters more of it.
The model is designed to predict cyclones with greater accuracy than existing systems. On average, the researchers say it gives forecasters a day more lead time, meaning its three-day forecasts match the accuracy that previous models achieved two days out.
A five-day signal before Hurricane Melissa
In October 2025, a storm system formed over the Caribbean Sea. Forecast models did not agree on what would happen next. One possibility was that the storm would stay weak and move toward Haiti. Another was that it would strengthen and head for Jamaica.
WeatherNext pointed to the second scenario. Five days before landfall, it predicted with 80 percent confidence that the system would strike Jamaica as a Category 5 hurricane.
Hurricane Melissa became catastrophic, bringing flooding and landslides across Jamaica. The source article says the model helped forecasters provide earlier warning to communities in the storm’s path, giving people and emergency systems more time to prepare.
That matters because hurricane readiness is not a single action. It can involve evacuations, positioning supplies, and moving response resources into place before conditions deteriorate. Mike Brennan, director of the US National Hurricane Center, summed up the stakes plainly: “Even a few hours can make a difference.”
Why intensity is so hard to forecast
Predicting a hurricane is not just about drawing a likely path on a map. Forecasters need to understand both track and intensity, and the source article makes clear that those are different technical problems.
Kate Musgrave, tropical cyclone group lead at the Cooperative Institute for Research in the Atmosphere and an author on the paper, explains that hurricanes operate across multiple spatial scales. Track forecasting depends on broad weather patterns, including cold fronts and prevailing winds. Intensity forecasting depends on smaller-scale local atmospheric and ocean conditions.
That difference has made intensity a persistent challenge. Earlier AI models performed well on storm track, Musgrave says, but not on intensity. The gap matters because a storm’s strength can change the level of danger dramatically.
Hurricane Melissa shows why. The source article says the storm system intensified rapidly and developed into an emergency situation overnight. It also says Melissa marked the first time the National Hurricane Centre was able to predict a Category 5 hurricane while the storm was only at a Category 1 stage.
How WeatherNext approaches uncertainty
Extreme events are difficult for artificial intelligence because machine learning depends on training data, and rare events do not provide much of it. Ferran Alet, a research scientist at Google DeepMind and one of the paper’s lead authors, describes the problem this way: “We don’t have that much cyclone data, but we have a lot of weather data.”
The team’s answer was to train a model that could handle both general weather and cyclones. Before WeatherNext entered live forecasting work, researchers tested it on retrospective data. Musgrave says the results were strong enough that researchers were skeptical the same performance would appear in real-time use. When forecasters began using it in operations, the performance held.
WeatherNext also does not return only one forecast. It produces a range of possible storm scenarios, helping forecasters account for the way small changes can grow into major differences later. The source article says the model created 50 scenarios per storm last year and now generates 1,000.
Forecasters can compare those outputs with other models as they judge how a storm may develop. Musgrave says that scale of scenario generation is not something existing numerical models can do with current computing power.
The black box question
One of the most striking details in the source article is that even the DeepMind researchers do not fully understand why the model performs as well as it does. WeatherNext uses much lower-resolution atmospheric data than traditional models typically require for intensity forecasting.
Alet says the reaction from the wider community was surprise because the lower-resolution inputs appear to contain more useful signal than previously believed. The model seems to be detecting something in that data that helps it anticipate storm intensity, but the researchers do not yet know exactly what.
That uncertainty cuts both ways. It may reveal a limit in explainability, but it may also point researchers toward new scientific questions about how cyclones work. Alet describes the model as a black box, while also saying it gives physicists a signal that something may be happening that was not previously understood.
A tool, not a replacement for forecasters
Brennan describes DeepMind’s model as a valuable addition to the forecasting toolbox, but he also cautions against treating any one model as definitive. A model that performs well in one season or for one storm may not necessarily be the best model for the next season or the next storm.
The human role remains central because a hurricane forecast is not only a prediction of track or intensity. Experts still have to translate those outputs into likely impacts for people on the ground. As Brennan says, “it's the impacts that kill people.”
Google DeepMind also announced that it is open-sourcing the WeatherNext models used during hurricane season. The goal is to let researchers use them, improve them, and potentially learn more about cyclone behavior.
For hurricane forecasting, the immediate promise is practical: more accurate warnings earlier in the timeline. For weather science, the deeper question is whether AI models like WeatherNext can help expose patterns that were previously hidden inside familiar data.