Introduction
Artificial intelligence is increasingly hailed as the future of weather forecasting—faster, cheaper, and often more accurate than traditional methods. Yet a surprising blind spot remains: when it comes to the most extreme and dangerous weather events, AI still stumbles. A new study published in Science reveals that leading AI models consistently underestimate record-breaking heat, cold, and wind, while older physics-based models still outperform them. This raises critical questions about how we prepare for and predict the very events that pose the greatest threat to society.
The Limitations of AI in Extreme Weather
Researchers led by Sebastian Engelke, a statistics professor at the University of Geneva, put some of the top AI weather models—including Google DeepMind's GraphCast and Huawei's Pangu-Weather—to the test against a database of recent extreme events. The results were clear: AI predictions frequently miss the mark for the most severe weather.
For example, during the 2020 Siberian heat wave—a record-breaking event that triggered widespread wildfires and permafrost melt—AI models consistently underestimated the high temperatures. (That heat wave was made 600 times more likely by climate change, according to a separate study.) Similarly, AI struggled to predict extreme wind speeds and record-breaking cold snaps.
Why AI Struggles with Extremes
The fundamental issue lies in how AI models are trained. 'They try to empirically understand, if I see a certain type of weather today, what is the weather tomorrow?' explains Engelke. 'Essentially, they are reproducing what has happened in the past.' Because extreme events are rare and often unprecedented—like a heat wave that breaks all previous records—the training data contains little to no information about such anomalies. The models simply have no prior examples to learn from.
Although some AI systems have since added probabilistic approaches to generate multiple possible outcomes, the underlying limitation persists. As Engelke notes, 'It's really the lack of information in their training data that makes it almost impossible for them to forecast it.'
Traditional Forecasting Still Leads for Extremes
Traditional physics-based weather models use complex mathematical equations to simulate the physical processes of the atmosphere. This approach does not rely on historical examples; it can adapt to new and unprecedented conditions more readily. While traditional models are far from perfect—they also struggle with extreme events—they still outperform AI when it comes to record-breaking weather, the study found.
'They do perform well on a lot of tasks, but for very extreme events—that are the most important for society—they still struggle,' says Engelke. The findings underscore that for the most dangerous events, we cannot yet rely solely on AI.
Where AI Excels: Typical Weather and Less Extreme Events
Despite its shortcomings at the extremes, AI shines in many other areas. For routine day-to-day forecasting, or for extreme weather that is not wildly outside historical bounds, AI models often outperform traditional ones. They can process vast amounts of data quickly and identify patterns that physics-based models might miss.
Consider the case of Storm Dennis, a rapidly intensifying cyclone that struck the UK in 2020. When Nvidia released its forecasting model Atlas earlier this year, it tested Atlas on Storm Dennis—even though the model had not been trained on that event. 'You can see just clearly by visualizing the magnitude of the wind and the magnitude of the pressure gradient that the model was able to capture realistically intense wind events and really intense cyclones that cause damage,' says Mike Pritchard, director of climate simulation research at Nvidia.
AI models are also adept at predicting hurricane paths. As a result, they are already being used alongside traditional models by weather agencies, data companies like The Weather Company, and in operational forecasting.
The Future of Forecasting: A Hybrid Approach
The study examined models from about a year ago, and AI weather forecasting continues to improve rapidly. Probabilistic versions that generate ensembles of forecasts are becoming more common. But as long as training data is the primary source of knowledge, AI will remain vulnerable to unprecedented extremes.
Experts suggest that the best path forward is a hybrid approach—combining the speed and pattern recognition of AI with the physical grounding of traditional models. This could help bridge the gap and deliver more reliable warnings for the weather events that matter most.
Conclusion
AI weather models are powerful tools, but they are not yet ready to replace traditional forecasting for the most dangerous extremes. The limitations highlighted by the Science study serve as an important reminder: as the climate continues to change and record-breaking events become more frequent, our forecasting methods must evolve to meet the challenge. For now, the old physics-based models still have a crucial role to play.