by Oliver Morrison, Pittsburgh’s Public Source
Pittsburgh’s Public Source is an independent nonprofit newsroom serving the Pittsburgh region. Sign up for its free newsletters.
Matthew Kramar has added a new step to his weather prediction routines: At the end of his analysis, he looks at a map of predictions generated by an AI model.
“I look at the AI and say – does it agree with me?” said Kramar, the science and operations officer for the National Weather Service in Pittsburgh. “If it does agree with me, then I feel pretty good about what my reasoning is. If it doesn’t, then I have to ask myself why.”
One of the reasons he’s added this extra step, he said, happened in May of 2024.
LISTEN to Oliver Morrison and The Allegheny Front’s Kara Holsopple discuss the story
Kramar said an AI weather model called Nadocast was predicting a roughly 3% to 5% chance of tornadoes southwest of Pittsburgh in two days. Although 3% may not seem very likely, Kramar said, it’s hundreds of times more likely than average. And the odds seemed to be increasing as the day got closer.
But when Kramar and his fellow meteorologists looked at the underlying weather data, he said, they didn’t see anything worrisome. “We were examining [the AI prediction] thinking, ‘Well this has got to be out to lunch,’” he said.
Then, on May 11, 2024, there were three confirmed tornadoes. The same thing happened a week later on May 17: The AI model showed a small chance of tornadoes and then six tornadoes appeared.
In surveillance footage from the 17th about two hours west of Pittsburgh in Carroll County, Ohio, a vortex appears out of seemingly nowhere and picks up a handful of docked pontoon boats and tosses them across the harbor. In hindsight, Kramar said, the weather service probably should have issued some kind of alert.
Then, on April 29 this year, “the atmospheric profiles were eerily similar” to May 2024, Kramar said.
This time, Kramar said, their understanding of the AI models had grown and they were prepared.
The challenge is that being prepared does not always mean knowing exactly what will happen, where it will happen or how disruptive a warning should be.
That tension was on display June 6, when country singer Morgan Wallen canceled his second Pittsburgh concert at Acrisure Stadium because of the threat of severe weather.
The National Weather Service eventually confirmed that six tornadoes landed that day, but they hit south of Pittsburgh and were long gone hours before the concert was scheduled to start.
The gap between a real regional threat and a clear local decision — whether to issue an alert, staff up or cancel a stadium concert — is exactly where meteorologists are trying to understand what AI can and cannot do.
The rapid rise of AI weather models
Using AI to help predict the weather is so new that one of the most important advances was developed by a single scientist on his laptop in 2022 while on sabbatical. Since then, AI weather models have proliferated and improved exponentially.
Universities and government agencies from across the United States, Europe and China have created weather-prediction models using AI, many of which they say are comparable to or better than more traditional models at predicting certain kinds of weather. Now some of the world’s biggest companies, including Nvidia, Google and Microsoft are competing to build even better weather models.
Andrew Parnell, a data science professor at University College Dublin, recently started a research center focused on how to use AI to predict the weather. His graduate students use AI models to look at how weather impacts things like air quality, strokes and infectious diseases.
“Weather forecasting has been completely transformed by our AI over the last two or three years,” he said.
But one of the biggest challenges in forecasting is predicting tornadoes. They form in the lower atmosphere where it’s harder for satellites and radar to measure what’s happening.
A single tornado event can cause more than $1 billion in damage; in 2025 alone six tornado events achieved this dubious distinction. A storm on May 31, 1985, that tore through Beaver County killed 89 people, the deadliest natural disaster in Pennsylvania since. One of the 43 tornadoes that day ripped a 750-foot-wide path of destruction through Beaver County, damaging 300 buildings across 39 miles, killing nine and injuring 120.
Paul Markowski, the head of the meteorology department at Penn State, was 10 years old at the time in central Pennsylvania. And news reports of what happened inspired him to become one of the country’s foremost tornado experts.
AI could help because the next generation of radar will produce more data than humans can reasonably process, Markowski said. Sophisticated “phased array” radar, which scans the sky more frequently, could generate terabytes of data nationwide every minute.
“There’s just no way a human can look at all of those scans,” he said.
AI could also help forecasters narrow who actually needs to be warned (and in what languages). Right now, about 80% of tornado warnings end up being false alarms for the people who receive them, Markowski said.
“You’re actually unnecessarily warning a lot of people,” he said.
That matters when storms threaten large crowds. “Let’s say you’ve got a football game here at Beaver Stadium. You’ve got 100,000 people sitting ducks. You’re going to need a longer lead time,” he said. “You could actually have things tuned in a way, where depending on what’s in the path, you could give different types of alerts.”
A complicated problem
Traditional weather models start by layering the globe with a mesh of small squares, each of which represents a mass of air that obeys the laws of physics. The models then use the air’s temperature, humidity and movement to calculate how the air pockets will crash into each other, fuse, transform and otherwise perform the complicated dance of fluid dynamics encoded by a string of math equations.
The addition of satellite data and improvements in radar technology in recent decades gave these models more precise starting points. Increases in computing power allowed forecasters to model weather on smaller squares, allowing for even more precise local predictions.
One of the reasons AI has made such rapid progress is that weather data is voluminous and consistent, according to Ryan Sobash, a scientist who studies severe weather at the National Center for Atmospheric Research, and works remotely from the Pittsburgh area. His decision to study meteorology at Penn State was seeded in part by the 1985 tornado, which damaged his family’s Beaver County property when he was a year old.
Severe thunderstorms are small and complex, and tiny differences in temperature, humidity or wind can lead to very different outcomes. To account for that uncertainty, forecasters often run models 100 times with slightly different starting points — a process that demands enormous computing power.
The development of GPU computer chips, Sobash said, has radically sped up this work. AI models mimic the underlying weather models by looking at the initial weather conditions and then guessing what will happen based on all of the past weather events it was trained on. The AI model doesn’t apply the laws of physics at all. Skipping over the physics allows the AI models to avoid the time-consuming math and predict 100 possible outcomes in minutes.
The AI model Sobash helped develop, for example, can predict thunderstorms three to eight days before they happen, with more accuracy than supercomputers, on a single laptop, in a fraction of the time of a traditional model.
A week before Wallen was supposed to take the stage, Sobash’s model predicted there would be storms near Pittsburgh and in two other parts of the country near Texas and the Dakotas. Seven days later, thunderstorms hit all three places.
“It’s a step change,” Sobash said.
Weather forecasting has been completely transformed by our AI over the last two or three years.
Andrew Parnell, professor of data science
If Wallen had looked at Sobash’s model, he would have had a heads up that a storm might be coming during his Pittsburgh tour stop, but it wouldn’t have given him enough precision to guarantee anyone’s safety that day.
Some researchers are using AI for shorter-term warnings. Jim Kurdzo, a senior researcher on MIT Lincoln Laboratory’s weather team, helped train an AI model on thousands of radar images of tornado-producing storms and storms that looked similar but never produced tornadoes.
The model could identify likely tornadoes 15 minutes before they happened better than traditional weather models, Kurdzo said. Meteorologists are still better than any single model, he said, but AI could help when forecasters are tracking several dangerous storms at once and they don’t notice a less obvious tornado.
But the model starts picking it up and it says, “‘Hey, I’ve got a detection here at 80%,’” which means the people in that tornado’s path also need to be sent a warning.
The future for AI weather models
Aaron Hill, an assistant professor at the University of Oklahoma, doesn’t think the AI models have enough of the underlying information that they need to predict tornadoes.
“The vast majority of these AI models are really large scale,” he said. “There’s not a lot of high-resolution, localized forecast models yet because it takes a lot of compute time and a lot of compute power to make those models. So, we don’t have something great for tornadoes” that will happen within a few hours.
Part of the problem, Hill said, is that forecasters still have limited data from the lowest part of the atmosphere: where the sun heats the Earth and makes the air “chaotic and turbulent,” like the air above a boiling pot. He and other researchers are testing specialized drones that could gather more of that lower-atmosphere data and help AI models create more hyperlocal forecasts.
Hill thinks tech companies like Google, Microsoft and Nvidia could make the fastest progress in this new era of AI weather prediction because they have so much computing power. These companies have been sharing some, if not all, of their research, he said.
Parnell, who leads an AI weather prediction group in Dublin, said the lack of complete transparency from private AI models threatens the open-data culture that weather science has long depended on. His group, for example, hasn’t been able to use Google’s most recent model because it’s not open-source.
“It’s frustrating,” Parnell said. “The reason why forecasting works so well in general is because everybody shares the data. I mean, we still share data with Russia.”
Google and Nvidia didn’t respond substantively to requests for comment. Microsoft declined an interview request but linked to its AI weather model Aurora, which is available to researchers and which at least one Carnegie Mellon professor has been experimenting with in his research.
The undiscovered tornado
The National Weather Service in Pittsburgh typically has two staff on a quiet day.
But there were six staff on hand on April 29, when the weather looked similar to that which led to eight tornadoes in Ohio two years ago.
Extra staffing gives them the ability to always have someone on the radar with their eye out for tornadoes, Kramar said, and another staffer ready to quickly communicate with local agencies if one shows up.
But on that day, the weather ended up being a couple of degrees cooler than predicted. And the twisters never came.
“What [AI] does is it helps heighten our awareness of, ‘Hey, there are environmental factors that might favor severe weather development,’” Kramar said. “And then we have to ask ourselves, ‘OK, great, those are there. Let’s anticipate that they’re possible. But let’s also look at how it can go wrong.’”
Oliver Morrison is the health and environment reporter at Pittsburgh’s Public Source and can be reached at ol****@**********ce.org.
This story was fact-checked by Katherine Weaver.
This article first appeared on Pittsburgh’s Public Source and is republished here under a Creative Commons Attribution-NoDerivatives 4.0 International License.
