Artificial Intelligence (AI) has rapidly transformed weather forecasting in recent years. Modern AI models deliver fast and energy-efficient predictions and, under average weather conditions, often achieve accuracy comparable to-or even exceeding-that of classical physics-based numerical models. However, for particularly severe, record-breaking extreme events, AI-based forecasts reach their limits. This is shown by a new international study led by Karlsruhe Institute of Technology (KIT) and University of Geneva. Results in Science Advances. (DOI: 10.1126/sciadv.aec1433 )
How well modern AI weather models predict extreme heat, cold, and wind events that exceed historical records was investigated by researchers led by Dr. Zhongwei Zhang at the Institute of Statistics at KIT. The result: under these exceptional weather conditions, the physics-based high-resolution model HRES of the European Centre for Medium-Range Weather Forecasts consistently outperforms the currently leading AI models.
AI Systematically Underestimates Records
The scientists compared several established AI models-including GraphCast, Pangu-Weather, and Fuxi-with the physics-based reference model HRES. While AI models perform well in overall evaluations across all weather situations, they show consistently larger forecast errors for record-breaking events. In particular, they underestimate both the intensity and the frequency of extreme events. "Our analyses show that AI models generally underestimate the intensity of heat, cold, and wind records," explains Zhongwei Zhang. "The greater the exceedance of the record of their training data, the larger the underestimation."
Limitations of Neural Networks underlying AI-Models
The researchers attribute this to a fundamental limitation of pure data-driven models: AI systems learn from historical data and are particularly effective at predicting weather patterns that resemble those observed in the past. By definition, however, record-breaking events lie outside previous experience.
"Neural networks struggle to reliably extrapolate beyond their training domain - that is, to make predictions beyond previously observed values," says Professor Sebastian Engelke, Full Professor at the University of Geneva and former supervisor of Zhongwei Zhang. "Physics-based models such as HRES, by contrast, are based on fundamental laws of physics. This ensures that their forecasts are still reliable when the atmosphere moves into states that have not yet been observed." Such record-breaking weather situations are occurring more frequently in a rapidly warming climate, with sometimes severe consequences for health, infrastructure, and the economy.
Implications for Early Warning Systems
The findings are particularly relevant for early warning systems and disaster management. A systematic underestimation of extreme events can result in warnings being issued too late-or not at all. The authors of the study therefore emphasize that AI weather models cannot currently replace classical numerical forecasts. "For high-risk applications, one should not rely solely on AI," states Zhongwei Zhang. Instead, the researchers recommend a parallel use of both approaches, as well as further research into hybrid models and physics-informed neural networks that combine physical knowledge with AI methods.
Perspectives for Improved AI Models
At the same time, the study identifies pathways for making AI-based weather forecasts more robust in the future. These include, among other measures, targeted enrichment of training data with simulated extreme events, new training methods from extreme value statistics, and hybrid modeling approaches. Until then, the central message remains: "AI is a powerful tool for weather forecasting-but for the most extreme and potentially high-impact events, physics-based models remain indispensable," Sebastian Engelke concludes.
Researchers from ETH Zurich, the Helmholtz Centre for Environmental Research, Technische Universität Dresden, and University of Geneva were also involved in the study.
Original publication:
Zhongwei Zhang, Erich Fischer, Jakob Zscheischler and Sebastian Engelke: Physics-based models outperform AI weather forecasts of record-breaking extremes. Science Advances, 2026. DOI: 10.1126/sciadv.aec1433 .
KIT Center Mathematics in the Natural, Engineering, and Economic Sciences: https://www.mathsee.kit.edu/
