Google DeepMind AI Model Reshapes Hurricane Forecasting

Isaias made landfall near Destin, Fla., as a Category 2 storm in early October. Conventional forecasting models initially predicted it would head toward southeastern Louisiana, but the Google DeepMind model called the western Florida Panhandle right out of the gate. Michael Lowry, a longtime Miami meteorologist and hurricane expert who has advised FEMA and the National Hurricane Center, said the model has become the standard. "Google DeepMind, in particular, has really proven itself the last few years as the gold standard model," Lowry said. "That model has been, by far and away, the best performing computer model that we've ever seen. Period."

The National Hurricane Center first incorporated the Google DeepMind model into its official forecasts last year. By the end of the 2025 Atlantic hurricane season, it was already one of the best at predicting a storm's path and strength, said Brian McNoldy, a hurricane researcher at the University of Miami. New models usually take years to refine after introduction. "It's something we're not used to – to have a model, the new kid on the block, very quickly becoming the best model out there," McNoldy said. "This pretty much came out of the gate in first place, and that's a pretty impressive feat."

Traditional physics-based models use real-time atmospheric data such as temperature, humidity, pressure and wind shear, measured by satellites and aircraft, and complex equations to predict a storm's path and strength. AI models rely more on history, identifying patterns by analyzing thousands of past storms and atmospheric data. "The AI models don't really know anything about the atmosphere other than the current picture," Lowry said. "Beyond that, it's not leaning on the governing equations of the atmosphere, the physics, to make predictions." Most people designing these models are not meteorologists but software engineers developing neural networks, Lowry said. A neural network is a set of computer algorithms modeled on the human brain and trained to find patterns in large amounts of data. That can make the models seem like a "black box," Lowry said: "We don't know why they might get a forecast right. We also don't know why they might get a forecast wrong."

McNoldy said relying on a system meteorologists do not fully understand can be unsettling, especially when millions of people may rely on forecasts during a storm. "It makes us a little uncomfortable, because we like to know why you're getting the right answer," he said. "And I feel like we don't – or at least I don't know why." Lowry said not all AI models have performed well and many remain unproven. The two most accurate so far, despite some pitfalls, are from Google DeepMind and the U.K.-based European Centre for Medium-Range Weather Forecasts, he said.

At the National Hurricane Center, forecasters use several proven models, both AI and traditional, when assembling official forecasts. "Every storm has different forecast challenges," NHC Director Michael Brennan said. "That's why we take that consensus approach and use that blend of the best models." Humans still interpret the model results before deciding what to tell the public. The NHC provides updates every six hours during active hurricanes, and models can sometimes show drastic changes from one update to the next. Forecasters must translate those shifting predictions into clear guidance. "That's one thing that the human forecaster provides a lot of, is continuity," Brennan said. "We're using our expertise and our understanding to translate that hurricane forecast into the hazard space, and that's what people really need to pay attention to."

Brennan said the Google DeepMind model still needs work. While it topped other models in tracking Isaias's path, it did not do as well as the NHC's blended forecast in determining the storm's intensity, or strength. Brennan's team gives regular feedback to Google DeepMind to help improve the model. Google did not respond to questions from NPR for the story. "The future of forecasting has always been changing," Brennan said.