Last fall, three days before Hurricane Melissa hit Jamaica, forecasters at the National Hurricane Center in Miami warned that the storm would strengthen rapidly from the least dangerous category to the deadliest. After that judgment turned out correct — among many accurate predictions the center was able to make during the unusually volatile storm — the forecasters attributed their success to artificial intelligence, which they incorporated earlier that year into their methodology.
Now, DeepMind, a Google unit that built the A.I. model, argues that the Melissa achievement was no fluke.
For a paper this month in Nature, Google DeepMind’s researchers analyzed hundreds of hurricane forecasts between 2023 and 2025. Their analysis demonstrates, they say, that their A.I. model can outdo other widely used prediction models by significant margins, effectively peering a day or more ahead of those other models.
“It’s a big deal,” said Amy McGovern, a professor of meteorology at the University of Oklahoma who directs an A.I. weather institute and played no role in the research. “If you can tell people a day in advance about a major hurricane, they’ll have more time to evacuate and prepare, in theory saving lives and property.”
Kerry Emanuel, an emeritus professor of atmospheric science at the Massachusetts Institute of Technology, welcomed the A.I. advance but cautioned that knowledgeable individuals still had to make sense of what the smart machines conjure up.
“It’s human beings who have to make these calls,” he said. Eventually, Dr. Emanuel added, “A.I. will be treated as just another form of guidance” along with satellite images and readings from hurricane hunter aircraft that pierce the tempests.
Hurricane Melissa, which made landfall on Oct. 28, was the strongest known hurricane to ever strike Jamaica, crushing homes with 185 mile-per-hour winds. Forty-five individuals lost their lives to the fierce storm — compared to the hundreds and thousands who perished in Caribbean hurricanes over the ages. The rise in forecast accuracy and safety precautions, experts say, has cut death tolls dramatically.
Melissa was a particularly tough case for hurricane forecasters because the storm’s behavior was so fickle. Many hurricanes move in a smooth arc. In contrast, Melissa made hairpin turns, sped up, slowed down and at one point even performed an enormous zigzag.
Adding to the challenge, the Caribbean waters that October were unusually warm, making them ideal for hurricane development. “We were on red alert,” recalled Philippe Papin, a National Hurricane Center forecaster who helped track Melissa.
The center’s forecasters say the A.I. model repeatedly aided them. For instance, it helped them predict Melissa’s true place of landfall when early forecasts swung widely among Haiti, Cuba and Jamaica.
The model also foresaw the storm’s quick strengthening. By definition, a hurricane that achieves sustained winds of 74 m.p.h. is known as a Category 1 storm, and of 157 m.p.h. or higher as a Category 5 — the most dangerous kind.
On Friday, Oct. 24, the model began predicting that Melissa would turn quickly into a Category 5 storm. Ferran Alet, a research scientist at Google DeepMind, recalled being “a bit nervous” because the team had never verified the model’s ability to make such a leap.
The next day, the National Hurricane Center echoed the edgy call. Never before had the center projected a Category 1-to-5 jump. “It was a first,” recalled Mr. Papin, noting that the call proved correct.
On Tuesday, Oct. 28, Melissa made landfall. The hurricane’s high winds sent roofs flying and turned rain into lashing torrents that caused wide flooding and landslides. Whole towns were transformed into rubble, leaving thousands of people homeless.
In March, the World Meteorological Organization decided to retire the name Melissa from the rotating list of hurricane names because the storm’s vast destructiveness quickly undid years and even decades of Jamaican development.
In a lengthy report in May on Melissa, the last superstorm of the 2025 Atlantic season, the National Hurricane Center praised the Google DeepMind model as having performed “remarkably well” in helping the center produce its early warnings.
In an interview, Daniel Brown, who heads the center’s hurricane unit, said that his forecasters had slowly gained confidence in the A.I. model’s performance over the course of the year and that their growing trust culminated in the case of Melissa.
Beginning last year, Google DeepMind made public an interactive visualization that lets users see how official hurricane forecasts, including those of the company’s model, measure up to real-life storms. Known as Google Weather Lab, the tool shows the development of the atmospheric disturbances in real time as well as in years past, as with Melissa.
The new Nature paper compares forecasts of the company’s model to official ones made for the global set of hurricanes that formed between 2023 and 2025.
The old forecasts included a type that experts consider the gold standard in weather forecasting — the Ensemble Prediction System of the European Center for Medium-Range Weather Forecasts, which is headquartered in Reading, England.
The DeepMind team compared such forecasts to those of its A.I. model on three of the most important features of any hurricane — path, intensity and structure or extent. A hurricane’s extent describes the total geographical area that will feel a storm’s impact, while intensity describes the peak winds at the storm’s center.
The team found that, on average, the A.I. model outdid the base line forecasts by a day or more on all three of the basic hurricane features. The model’s five-day forecasts for storm paths, for example, achieved an accuracy lead time of more than 30 hours over its best public rival.
“Lead time is one of the most important things,” said Mike Brennan, director of the National Hurricane Center. “If you’re short on lead time, there may not be enough time to evacuate the hospital or shut down the power facility.”
While giving DeepMind credit for the step forward, Dr. McGovern of the University of Oklahoma noted that it would not have been possible without the National Hurricane Center’s input.
DeepMind’s experts “worked hand in hand with their partner, the government, and that’s what made it a success,” she said. “It was a cooperative development.”
