Hurricanes and AI: Forecasting Takes a Major Step Forward with Google DeepMind

Hurricane Melissa was barely a tropical depression — winds under 39 miles per hour — when it entered the Caribbean in October 2025. Four days later it made landfall on Jamaica as a Category 5 monster with 185-mile-per-hour winds, one of the strongest storms on record to strike the island Google DeepMind Blog.

What almost nobody noticed at the time was that one model saw it coming clearly and early. Google's DeepMind WeatherNext system had predicted Category 5 intensity with 80 percent confidence five days before landfall, climbing to near certainty three days out. The National Hurricane Center had never successfully made that kind of call from such a modest starting point — a storm barely organized enough to earn a name — before.

That result is now on the record. The NHC's end-of-season verification report confirms that DeepMind's ensemble was among the top-performing models for both hurricane track and intensity across the full 2025 Atlantic season — the first time AI forecasting was evaluated alongside the physics-based models that have defined operational meteorology for decades NHC Verification Report 2025. Former NHC branch chief James Franklin, who conducted an independent analysis of the season's model performance, found that Google's ensemble beat every other computer model guidance on both metrics for the year CNN.

The 2025 season ended with three Category 5 landfalls — the most since 2005 Tallahassee Democrat. And the AI did not miss any of them.

"WeatherNext was a step change," said Dr. Wallace Hogsett, science and operations officer at the NHC, speaking at the American Meteorological Society's tropical meteorology conference in April WFLA. "It gave us confidence in situations where we would not have had it before."

The model runs in minutes on cloud hardware rather than hours on a supercomputer, can produce ensembles of fifty or more scenario variations in the time a traditional model takes to complete a single run, and does not require solving the Navier-Stokes equations that underpin every physics-based forecast. It learned the atmosphere the way vision systems learned faces: by watching an enormous number of examples and finding the patterns that matter Google DeepMind WeatherNext 2.

AccuWeather senior hurricane forecaster Alex DaSilva, whose firm conducted its own analysis of the 2025 season, put it more bluntly: "They just built this model a few years ago and it's already the highest-performing model, even beating the National Hurricane Center in some forecasts" Tallahassee Democrat.

But the same qualities that make WeatherNext powerful are also what make some meteorologists uneasy.

"It's a black box," said James Franklin. "It looks for patterns in past data. It's not tied to physics. It can come up with any answer that it thinks it finds in the data" CNN.

That distinction matters more than it might appear. A model that has never seen a hurricane behave in a genuinely unprecedented way — one shaped by conditions that fall outside the historical distribution it was trained on — could produce confident, catastrophically wrong forecasts with no signal that anything has gone wrong inside the system. Meteorologists can reason backward from physical first principles when a model goes off the rails. They cannot yet do that with a neural network.

This is not a hypothetical concern. NOAA's own first-generation AI global forecast system, AIGFS, which uses a Google DeepMind architecture as its foundation, showed measurable degradation on tropical cyclone intensity forecasts in its initial version NOAA Weather.gov. WeatherNext appears to have avoided that specific problem through its specialized training on extreme tropical cyclone datasets, but the episode illustrates a broader truth: AI models can fail in ways that are easy to miss until something goes wrong at scale.

The insurance industry is watching closely. After a decade in which catastrophe models underpriced hurricane risk — and after a year in which three Category 5 landfalls rewrote the record books — the question for reinsurers and primary insurers is whether AI's improved intensity forecasts can be integrated into existing risk frameworks fast enough to matter for the 2026 season. If WeatherNext's five-day lead time on Category 5 calls becomes reliable, it changes the economics of evacuation orders, the pricing of coastal property insurance, and the actuarial assumptions underlying trillions of dollars in exposure. Nobody has figured out yet how to put a number on "the model said Category 5 five days out and it was right."

The NHC's official position is appropriately hedged. Hogsett described AI as "a component" of the hurricane forecast process rather than its foundation. The existing ensemble of physics-based models, including the European Centre for Medium-Range Weather Forecasts model and the U.S. Global Forecast System, remains the backbone of official NHC guidance. WeatherNext supplements them, not replaces them — at least for now NOAA Weather.gov.

What changed in 2025 was that the NHC started treating AI as something close to an operational tool rather than an experiment. The 2026 Atlantic hurricane season, which began June 1, is the first full season in which AI forecasting systems are a routine part of the morning briefing rather than a novelty. WeatherNext 2, the latest generation of the model released by Google DeepMind and Google Research in November 2025, has been integrated into Google Search, Gemini, Pixel Weather, and the Google Maps Platform Weather API — meaning hundreds of millions of people are already receiving AI-modified weather forecasts without necessarily knowing it Google DeepMind WeatherNext 2.

The question the 2026 season will start to answer is not whether AI can outperform physics-based models on a good day. It demonstrably can, and did. The question is whether the trust gap can be closed fast enough for the technology to fulfill what its proponents claim for it: a fundamental improvement in the ability of coastal communities to prepare for the storms that are, because of climate change, growing both stronger and harder to forecast through traditional methods alone.

The black-box problem is real. So is the capability. Both things are true at the same time, and nobody at the NHC or anywhere else has figured out yet how to be equally confident in both.