Forecasting the Unforecastable
Day-to-day weather forecasts have become remarkably accurate, but predicting rare, extreme events—such as the deadliest heat waves—remains a formidable challenge. These events, sometimes called 'gray swans,' occur so infrequently that they fall outside the range of typical weather patterns. Traditional supercomputer-based models can simulate them, but they demand enormous computational resources and time. On the other hand, newer AI-based forecasting models excel at routine predictions but often fail to capture outliers that are underrepresented in their training data.
This gap has motivated researchers to seek a middle ground. A team led by scientists at the University of Chicago and collaborators in France has developed a hybrid method that combines the strengths of both approaches. Published in the journal Physical Review Letters, the new technique promises to improve the prediction of rare, high-impact weather events while using far fewer resources than traditional models.
The Challenge of Gray Swans
Heat waves are among the deadliest forms of extreme weather. In 2003, a heat wave across Europe caused an estimated 70,000 deaths, and in 2010, Russia experienced a heat wave that led to 56,000 deaths. More recently, in June 2026, nearly half of the United States—about 180 million people—experienced dangerous temperatures. These events are becoming more frequent and severe due to climate change, making accurate forecasting ever more critical.
However, the very nature of such outliers makes them difficult to study. They are rare, so there is limited historical data to train machine learning models. Traditional physics-based models, which simulate the atmosphere's behavior from fundamental equations, can generate extreme scenarios, but they require massive computing power and can take days to run. This makes them impractical for real-time forecasting or for exploring many possible scenarios.
As Pedram Hassanzadeh, associate professor of geophysical sciences at the University of Chicago and co-leader of the research, explains, 'AI weather and climate models are one of the great achievements of AI in science, but they're not magical—they fail on gray swans, the rarest and most extreme events. Detailed physics-based models can capture extremes, but they require prohibitively large amounts of time and energy.'
A Hybrid Solution
The new method, developed by Hassanzadeh's Climate Extremes Theory and Data Group and international collaborators, integrates AI with traditional physics-based modeling. The approach leverages the speed of AI to explore a wide range of possible conditions, while using physics-based models to ensure that the rare events are physically realistic. This combination allows the team to estimate the probability of extreme events more accurately and efficiently.
The key innovation lies in how the two components interact. The AI model, trained on a limited set of extreme event simulations, proposes candidate scenarios. These are then validated and refined using the physics-based model, which corrects any unrealistic outcomes. This iterative process converges on a reliable probability distribution for rare events, even when the AI alone would miss them.
'The power of this method is that it combines the strengths of both AI and traditional physics and is particularly effective for extreme events, which are the hardest to simulate and have the greatest societal impact,' said Hassanzadeh.
Implications for Forecasting and Climate Research
The hybrid approach could significantly improve our ability to anticipate rare but devastating weather events. For example, it could help emergency planners prepare for heat waves, floods, or droughts that might occur once in a century or even a millennium. By providing more accurate probabilities, the method could inform infrastructure design, public health strategies, and disaster response plans.
Moreover, the reduced computational cost means that such forecasts could be run more frequently and for longer time horizons. This could lead to better seasonal and decadal predictions, helping societies adapt to a changing climate.
The research also underscores the importance of combining AI with physical understanding. While AI has revolutionized many fields, it has limitations when it comes to rare events. By blending the two, scientists can harness the best of both worlds.
Looking Ahead
The team plans to refine the method further and test it on other types of extreme events, such as hurricanes and tornadoes. They also aim to make the tool accessible to weather services and climate researchers worldwide.
As extreme weather becomes more common, the need for accurate forecasting grows. This hybrid approach offers a promising path forward, one that could save lives and reduce economic losses. By improving our ability to predict the unpredictable, we can better prepare for the challenges that lie ahead.
For now, the method represents a significant step forward in the science of extreme event prediction. It is a testament to the power of interdisciplinary collaboration and the potential of combining cutting-edge AI with time-tested physical principles.
This article is based on reporting by Phys.org. Read the original article.
Originally published on phys.org
