DeepMind hurricane forecasting model incorporated into official forecasts by the US National Hurricane Center

In early October, Hurricane Isaias made landfall near Destin, Florida, at Category 2 intensity. Conventional forecast models initially judged that it would continue westward and head toward southeastern Louisiana. But Google DeepMind's artificial intelligence model got it right from the start: the hurricane would strike the western part of the Florida Panhandle.

"Google DeepMind in particular has really proven itself over the past few years to be the gold-standard model," said Michael Lowry, a veteran Miami meteorologist and hurricane expert who has advised the U.S. Federal Emergency Management Agency and the National Hurricane Center. "That model is the best-performing computer model we've ever seen, by far. That's just how it is."

The National Hurricane Center began including the Google DeepMind model in its official forecasts for the first time last year. By the end of the 2025 Atlantic hurricane season, it was already one of the best models for predicting storm tracks and intensity, said Brian McNoldy, a hurricane researcher at the University of Miami. Normally, new models need years of refinement after being introduced. "This is a situation we're not used to — a new model that comes out of nowhere and quickly becomes the best model," McNoldy said. "It ranked No. 1 almost as soon as it debuted, which is a pretty remarkable achievement."

Traditional physics-based hurricane models rely on atmospheric data measured in real time by satellites and aircraft, such as temperature, humidity, pressure and wind shear, and then use complex equations to calculate a storm's likely track and intensity. Artificial intelligence models work in a completely different way: they rely more on history, identifying patterns by analyzing thousands of past hurricanes and the accompanying atmospheric data.

"AI models actually know nothing about the atmosphere beyond the current picture," Lowry said. "Beyond that, they don't rely on the governing equations of atmospheric motion — that is, the laws of physics — to make predictions." Lowry noted that most of the people who design and develop these models are not even meteorologists: "They're software engineers, doing neural network development." So-called "neural networks" are sets of computer algorithms modeled on the human brain that are trained to find patterns in massive amounts of data.

As a result, these models look like a "black box." Lowry said: "We don't know why they forecast correctly, and we don't know why they forecast incorrectly." For meteorologists, relying on a system they don't fully understand can be unsettling, especially during a storm on which millions of people may be depending for forecasts. "It makes us a little uncomfortable, because we want to know why you're arriving at the right answer," McNoldy said. "And I think we don't know — at least I don't."

Still, not all AI models perform well. Lowry said there are a large number of untested AI weather forecasting models on the market. The two currently proven to be the most accurate, despite each having flaws, are the models developed by Google DeepMind and by the UK-headquartered European Centre for Medium-Range Weather Forecasts.

In forecasting the track of Isaias, Google DeepMind's AI model beat the traditional physics-based models, but it did not do a good job of predicting the storm's intensity. A photo taken by the Associated Press on October 10 shows people in Destin retrieving items from a collapsed boathouse after Isaias.

Forecasters at the National Hurricane Center use multiple validated models when producing official forecasts, both AI models and traditional models. "Every storm has a different forecasting challenge," said Michael Brennan, director of the National Hurricane Center. "That's why we take an ensemble approach and use a mix of the best models." And it is still up to humans to interpret the results of these models and then decide what to tell the public. During active hurricanes, the National Hurricane Center issues updates every six hours. The models themselves can sometimes change dramatically between updates, and forecasters must turn these fluctuating predictions into clear guidance that people can act on.

"A big part of the value that human forecasters provide is continuity," Brennan said. "We use our own expertise and understanding to translate hurricane forecasts into hazard information, and that's what people really need to pay attention to."

Scientists expect the technology to advance rapidly. Brennan said the Google DeepMind model still needs improvement. It outperformed other models in tracking the path of Isaias, but it was not as good as the National Hurricane Center's blended forecast at judging storm intensity. Brennan's team regularly provides feedback to Google DeepMind to help improve the model. Google did not respond to NPR's questions on the matter.

Brennan said the "future of forecasting is always changing," and "it's never just one thing driving these advances." He noted that not long ago forecasters could not predict how strong a storm would be within two days; that changed because of advances in forecast models and better data collected by satellites and aircraft. "For intensity forecasts, the error is about half what it was around the time of Hurricane Katrina," Brennan said.

Meteorologist Lowry believes what is exciting about AI is the speed of improvement. AI models can be updated in months or even weeks, while traditional models are updated at most once a year. "The skill and hurricane forecasting progress they achieve in a year is what it used to take us 10 years, or even more than a decade, to accomplish," Lowry said.