AI hurricane models forecast tracks well but struggle with storm intensity

The National Hurricane Center began using new AI weather prediction systems as guidance during the 2025 hurricane season, after spending 2024 and 2025 testing and verifying the tools alongside traditional forecasts. Wallace Hogsett, the center's Science Operations Officer, said the pace of new AI tools emerging was faster than at any point in his career. NHC has partnered directly with Google DeepMind on a hurricane forecast model, and has also evaluated AI-driven global systems built by NOAA's Environmental Modeling Center and the European Center for Medium-Range Weather Forecasts.

One storm stood out during that first operational season. Hurricane Melissa was, in Hogsett's words, a very difficult-to-forecast and impactful storm, and the AI models honed in early on its likely track and intensity, complementing the traditional physics-based guidance. Google DeepMind's own account credits its model with predicting Melissa's rapid intensification and landfall in Jamaica, giving the Hurricane Center's forecasters advance warning and giving teams on the ground more time to prepare. Hogsett cautioned that this is just one example, and that in other cases the traditional models performed better.

Traditional weather models start from the current state of the atmosphere and solve the physical equations that govern it, a process that requires large supercomputers to run for each forecast. AI weather prediction models start from the same kind of starting data, but instead of solving those equations, they draw on patterns learned by studying decades of historical atmospheric data, according to Hogsett. During training, the AI systems learn how variables like pressure, wind, and temperature relate to each other over space and time, then use those learned relationships to estimate what comes next. Because the relationships are learned in advance rather than recalculated each time, AI models run far faster than traditional ones.

Google DeepMind's WeatherNext Cyclones model, described in a paper published in Nature, was trained end-to-end on nearly 20 terabytes of global atmospheric data together with a historical record of storm observations spanning nearly 5,000 storms. Benchmarked against historical cyclones from 2023 and 2024, the model gained more than a full day of lead time over other top weather models for track, intensity, and wind structure forecasts. Google DeepMind describes this jump as equivalent to roughly a decade of typical meteorological progress. The model can now generate a 15-day forecast in under a minute on specialized computer chips called TPUs, and this year it is producing 1,000 possible storm scenarios at once, up from 50 the year before, to help capture rare but consequential outcomes like sudden strengthening.

A study from Rice University, published in the Journal of Geophysical Research: Atmospheres, took a different approach: rather than looking at headline cases, researchers simulated roughly 200 storms from the North Atlantic and western North Pacific between 2020 and 2025, deliberately choosing storms outside the periods the AI models were trained on. They tested two AI global weather models, Pangu-Weather and Aurora, comparing their storm output against a reference dataset called ERA5 that blends historical observations into a consistent record. Storm tracks came out well: the models reproduced where storms traveled and where they made landfall with a high degree of consistency, according to Rice civil and environmental engineering professor Avantika Gori, who led the study. Intensity was a mixed picture. Earlier AI systems tended to underestimate how strong storms got, and in the Rice tests, Aurora matched the reference data's intensity levels more closely, while Pangu-Weather showed larger errors for the strongest cyclones. The researchers noted a caveat that cuts both ways: the reference dataset itself tends to underestimate peak storm intensity compared with real-world observations, so matching it does not guarantee the models are matching reality.

The more significant finding concerned the internal structure of the simulated storms, meaning how their winds are arranged around the center rather than simply how strong they are. Wind patterns that looked realistic on screen did not always hold up against gradient wind balance, a physical relationship that governs how winds and pressure differences should relate to each other in a mature cyclone. The researchers found notable deviations from that relationship near storm centers. Both AI models also tended to overestimate the size of the storm's inner core, particularly in stronger storms. Because a storm's wind damage, rainfall, and storm surge depend on how its winds are organized and not just its overall category, Gori said these structural inconsistencies can carry through into hazard predictions that emergency planners rely on.

Separately, a 2023 study published in Weather and Forecasting and archived by NOAA developed a machine learning system aimed specifically at predicting rapid intensification, when a storm's winds strengthen unusually fast. Built using data from the operational Hurricane Weather Research and Forecasting model, the system reached a 56% detection rate for rapid intensification with a 46% false alarm rate, compared with 10 to 30% detection and 50 to 60% false alarms for the operational models it was tested against. The study also found that a storm's inner-core humidity was an especially useful signal for anticipating this kind of rapid strengthening.

Both the Hurricane Center and the Rice researchers frame AI as a complement to existing tools rather than a substitute for them. Hogsett said that despite the number of new models emerging, none of them are perfect, and the Hurricane Center's own combined forecast, produced by human forecasters weighing multiple models, has remained more skillful and consistent than any individual model across many storms and seasons. He said AI systems are also harder to interrogate than traditional models: because they generate answers through learned patterns rather than explicit physical equations, researchers are still working to understand not just what a model predicts but why. Gori raised a related concern, noting that because these models have millions or billions of internal parameters, researchers don't always have visibility into how they generate their predictions, which she said makes systematic evaluation especially important for high-consequence events like tropical cyclones.

Asked directly whether AI will eventually replace human hurricane forecasters, Hogsett said the answer, based on what NHC has learned so far, is a resounding no. He argued that trusted experts remain necessary to observe, synthesize, and make sense of the growing volume of model output, and to communicate a coherent risk message to the public. Gori reached a similar conclusion from the research side, saying AI weather tools are extraordinarily powerful but are not self-validating, and that close collaboration between atmospheric scientists and AI developers will be needed to keep model outputs physically meaningful as the technology develops further.