Michael Lowry, a Miami-based meteorologist and hurricane specialist who has advised federal emergency management and forecasting agencies, described Google DeepMind's model as a standout performer.
“That model has been, by far and away, the best performing computer model that we've ever seen. Period,” Lowry said.
The model's rapid improvement has surprised researchers accustomed to traditional forecasting systems, which often require years of development and refinement. Brian McNoldy, a hurricane researcher at the University of Miami, said DeepMind's system was already delivering leading results for storm trajectories and intensity by the conclusion of the 2025 Atlantic hurricane season.
“This pretty much came out of the gate in first place, and that's a pretty impressive feat,” McNoldy said.
DeepMind's approach differs substantially from the physics-based forecasting methods that meteorologists have used for decades. Conventional systems process atmospheric measurements collected by satellites and aircraft, including wind conditions, temperature, humidity and pressure. They then apply mathematical models of atmospheric behavior to estimate how storms will develop and move.
AI forecasting systems instead learn relationships from historical weather records and thousands of previous storms. Using the current state of the atmosphere as their starting point, they generate predictions based on patterns identified during training rather than explicitly calculating the physical processes behind every change.
That approach has delivered impressive results, but it also introduces challenges for meteorologists who need to understand why a forecast is reliable.
“We don't know why they might get a forecast right. We also don't know why they might get a forecast wrong,” Lowry said.
McNoldy similarly acknowledged that the limited visibility into AI-generated predictions can make researchers uneasy. Traditional models are built around established physical principles, while neural networks can produce accurate results without offering an equally clear explanation of how they reached them.
Not every AI weather system has demonstrated comparable reliability. Researchers identify Google DeepMind and the European Centre for Medium-Range Weather Forecasts (ECMWF) as the developers of the two strongest-performing AI models so far, although both still have limitations.
Hurricane Isaias illustrated those differences. DeepMind's model performed particularly well in identifying the storm's eventual path, but its prediction of the hurricane's intensity was less accurate than the National Hurricane Center's combined forecast.
That distinction helps explain why federal forecasters continue to rely on multiple systems rather than selecting a single AI model.
The NHC combines predictions from established AI and conventional models when preparing its official hurricane forecasts. Director Michael Brennan said different storms present different forecasting difficulties, making a consensus approach more dependable than relying entirely on one system.
“Every storm has different forecast challenges,” Brennan said. “That's why we take that consensus approach and use that blend of the best models.”
Human meteorologists remain responsible for evaluating the predictions and communicating potential hazards. During active hurricanes, the NHC publishes updated forecasts every six hours, and individual computer models can produce substantial changes between successive runs.
Forecasters use their experience to assess those fluctuations and provide consistent guidance about the risks facing affected communities. Brennan emphasized that translating technical predictions into practical warnings remains a central part of their work.
AI is contributing to improvements in a field that has already benefited from advances in observational technology and numerical forecasting. According to Brennan, errors in predicting hurricane intensity have fallen to roughly half the levels recorded around the time of Hurricane Katrina.
Researchers also expect AI systems to improve more rapidly than conventional models. Lowry said traditional forecasting software generally receives major updates no more than once a year, while AI models can be refined within weeks or months.
He suggested that recent advances in AI forecasting accuracy have compressed progress that previously might have required a decade into a much shorter development cycle.
The National Hurricane Center is also providing feedback to Google DeepMind to help address weaknesses in its predictions, including the accuracy of storm-intensity estimates.
Despite the technology's rapid development, Lowry does not see AI as a replacement for professional meteorologists. Instead, he considers it an additional forecasting resource that can strengthen the information available to human experts.
“If anything, I think it is a boon for our business,” he said.
DeepMind's successful prediction of Hurricane Isaias demonstrates how quickly AI weather models have become useful in operational forecasting. Their continued adoption, however, depends on balancing improved predictive accuracy with the judgment of meteorologists who must interpret the results and communicate risks to the public.
This analysis is based on reporting from NPR.
Image courtesy of www.polarwx.com/tropical.
This article was generated with AI assistance and reviewed for accuracy and quality.
