Chicago, IL and New York, NY, April 08, 2026 (GLOBE NEWSWIRE) -- At a time when artificial intelligence (AI) is revolutionizing weather prediction with its ability to analyze massive datasets and deliver faster, more accurate daily forecasts, researchers from the University of Chicago and New York University and their collaborators have found a critical gap in its predictive power: rare, extreme weather events. They have also made progress toward offering solutions.
Examples of these extreme events include fierce tropical cyclones that are scarcer than even category 5 cyclones such as Hurricane Harvey, which made landfall in Texas in 2017 and became one of the costliest and most destructive hurricanes in U.S. history. Working with researchers at the University of California, Santa Cruz, the team identified a major weakness in current AI systems: they often miss or downplay “gray swan” storms. These events are considered low‑probability and treated as outliers, so pattern‑based algorithms tend to dismiss them. A gray swan is a known risk that goes under‑prepared for, turning an “unlikely” event into a crisis with severe, far‑reaching consequences. As AI systems are rapidly adopted by weather companies and national meteorological agencies, the risk of such failures is growing.
The good news, however, is that there’s a way around it. The team, along with collaborators from École Normale Supérieure in Paris, have shown that by combining AI with traditional physics-based predictive models through mathematical algorithms, the reliability of current weather models is significantly boosted and these extreme gray swan events can in fact be forecasted.
A summary of these findings is presented in a recent article in SIAM News, a publication of Society for Industrial and Applied Mathematics (SIAM).
“This discovery is significant because climate change is increasing the odds of these high-impact extremes happening, posing potentially serious threats that demand better preparation,” said Pedram Hassanzadeh, associate professor of Geophysical Sciences and Computational and Applied Mathematics at the University of Chicago and Chair of the SIAM Activity Group on Mathematics of Planet Earth.
Hassanzadeh co-authored the paper, entitled Forecasting the Unseen: AI Weather Models and Gray Swan Extreme Events, with Y. Qiang Sun, Research Scientist at the University of Chicago, Jonathan Weare professor of mathematics in the Courant Institute of Mathematical Sciences at New York University, and Dorian S. Abbot, professor of geophysical sciences at the University of Chicago.
“While AI weather models are good and one of the biggest achievements in AI in science, they’re not magical,” emphasized Hassanzadeh, whose team explored a range of popular AI weather models, including NVIDIA’s FourCastNet, Google DeepMind’s GraphCast and GenCast and Microsoft's Aurora.
“Because they learn from patterns in historical data – typically covering only about the last 40 years – they struggle to produce events they’ve never seen before,” he said. “So, if an AI model has never seen a gray swan event, it is unlikely to predict one reliably,” he said.
That’s where mathematical modeling and physics-based simulations come in. By building a feedback loop between AI and physics-based models through a mathematical algorithm, they can build a forecasting framework that can make better weather predictions, especially for rare events, explained Sun.
The team has found that exposure to even a small number of modeled extreme events — as few as five cases — can substantially improve an AI model’s ability to recognize and predict gray swan behavior.
Interestingly, their research also shows that AI, mathematics, or physics alone are not complete solutions. “Traditional physics models on their own are too expensive and time-consuming while AI methods on their own aren’t enough for extremes,” Hassanzadeh said. “The power of mathematics has made it possible to combine the two, enabling us to study very strong events that we’ve never seen before.”
This includes the potential to capture extreme weather events that will happen 20 or 30 years from now as the climate warms, he explained. “These events might occur once in 100, 1,000 or 10,000 years, but once they do, they will have significant societal impact,” he said, emphasizing that current AI models will have challenges with such events. The floods caused by Hurricane Harvey, for example, were considered a once-in-a-2,000-year event.
At the same time, a surprising result the team found is that AI models on their own can forecast events based on learnings from other regions, which was not something that was built into them. “This is an encouraging finding that was not expected,” Hassanzadeh said.
“It means that the AI models can forecast an event that had no precedent in one region, but occurred once in a while in another region,” he explained. For example, they were able to use tropical cyclone patterns from one ocean basin, such as the Atlantic, to predict tropical cyclones in other ocean basins, such as the Pacific. They reached the same conclusion by examining the unprecedented 2024 rainfall over Dubai.
According to Sun, developing a better understanding of what current AI weather models can and cannot do will ultimately lead to the building of better and more reliable weather systems.
“AI weather models are revolutionizing forecasting techniques, but scientists don’t yet fully understand their limits, and how and what they learn,” Sun explained. “To predict what AI has not seen before requires increased collaboration between atmospheric scientists, mathematicians, and computer scientists.”
About Society for Industrial and Applied Mathematics (www.siam.org)
Society for Industrial and Applied Mathematics (SIAM), headquartered in Philadelphia, Pennsylvania, is an international society of 14,000 individual, academic, and corporate members from 85 countries. SIAM fosters the development of applied mathematics and computational methodologies needed in various application areas. Through publications, conferences, and communities like student chapters, geographic sections, and activity groups, SIAM builds cooperation between mathematics and the worlds of science and technology to solve real-world problems. Learn more at siam.org.
