UM Study: AI Could Improve Hurricane Forecasting, With Limits

AI beat traditional hurricane forecasting models at spotting storms days out. Closer to formation, the AI actually did worse.

August 28, 2026
UM Study: AI Could Improve Hurricane Forecasting, With Limits Miami

Summary: A University of Miami study led by atmospheric sciences professor Sharan Majumdar found that ECMWF's AI Forecasting System produced higher tropical cyclone development probabilities than its traditional model at lead times of roughly 84 to 120 hours, particularly for stronger tropical waves, but produced lower probabilities than the traditional system at shorter lead times of 36 to 48 hours, especially for weaker systems. The AI system also showed a consistent advantage in forecasting storm location, though Majumdar was clear the findings support AI as a complement to traditional hurricane forecasting rather than a replacement for it.

hurricaneforecastinguniversityofmiamiaiweatherecmwfatlantichurricanesclimatescience

AI beat traditional hurricane forecasting models at spotting storms days out. Closer to formation, the AI actually did worse.

That's the core finding from a new University of Miami study. Researchers compared ECMWF's traditional forecasting system against its newer AI-based one.

What the Study Actually Found

The study was led by Rosenstiel School professor Sharan Majumdar, based at the University of Miami's Coral Gables campus.

It examined African easterly waves and tropical cyclone development across the Atlantic from 2020 through 2024. These waves are the disturbances that often organize into hurricanes.

Among 18 tropical cyclones that developed in 2024, the AI system frequently produced higher development probabilities than the traditional model. That held true at lead times of roughly 84 to 120 hours, particularly for stronger tropical waves.

The AI system also showed a real advantage in one specific area. Its average position error for developing storms was often smaller than three other forecasting approaches tested. That included the traditional model's own ensemble.

Where AI Fell Short

The advantage flipped at shorter lead times.

Between 36 and 48 hours before formation, the AI system generally produced lower probabilities than the traditional model. That gap was especially pronounced for weaker systems.

That's a meaningful limitation, not just a footnote. The traditional model has consistently improved over the study period, including a 2023 upgrade that cut its grid spacing in half.

The gap between the two systems points to something forecasters already know. The signal that a disturbance will become a named storm can shift substantially as the event approaches.

Why This Isn't About Replacing Forecasters

Majumdar was direct about what the findings don't support. "The findings do not suggest that AI should replace traditional numerical weather prediction," he said.

Instead, he described AI forecasts as a complement, giving forecasters another data source when assessing storm potential. That's a meaningfully different claim than "AI forecasting works better."

"Better identification of developing systems several days in advance could give forecasters more time to monitor disturbances and assess possible tracks," Majumdar said.

ECMWF's AI system became fully operational in 2025.

The study was funded by the National Science Foundation and ECMWF. Collaborators included researchers from the University of Bonn and NSF's National Center for Atmospheric Research.

Why South Florida Should Care

This is genuinely important research for a region that lives with hurricane risk every season. Extra lead time on storm formation, even a partial improvement, translates directly into more preparation time for South Florida communities.

The mixed results matter just as much as the improvements. Residents shouldn't expect AI to replace the National Hurricane Center's traditional forecasting anytime soon. A University of Miami study explains exactly why not yet.

Inside How AI Robots Actually Get Debugged in Production