KIT and University of Geneva Research Reveals Key Limitations of AI Forecasting in Predicting Record-Breaking Weather Phenomena

Temperature anomalies during the 2020 heat wave in Siberia, which broke historical records and caused severe wildfires, among other things. (Image: Zhongwei Zhang, KIT)

(IN BRIEF) A new international study led by Karlsruhe Institute of Technology and the University of Geneva finds that while AI-based weather forecasting models are highly effective under typical conditions, they fall short when predicting record-breaking extreme weather events such as intense heat waves, cold spells, and strong winds. The research shows that AI systems, including GraphCast, Pangu-Weather, and Fuxi, tend to systematically underestimate both the severity and frequency of extreme events, largely due to their reliance on historical training data. In contrast, the physics-based HRES model from the European Centre for Medium-Range Weather Forecasts demonstrates greater reliability in these scenarios because it is grounded in fundamental atmospheric laws. The findings highlight important implications for early warning systems and disaster response, suggesting that exclusive reliance on AI could lead to underpreparedness. The study recommends combining AI and traditional models while advancing hybrid approaches to improve forecasting accuracy in a changing climate where extreme events are becoming more frequent.

(PRESS RELEASE) KARLSRUHE, 4-May-2026 — /EuropaWire/ — The Karlsruhe Institute of Technology and the University of Geneva have led an international research effort revealing that artificial intelligence-driven weather forecasting systems face significant limitations when predicting record-breaking extreme weather events. The findings, published in Science Advances, provide new insights into the comparative performance of AI-based and traditional physics-driven forecasting approaches.

In recent years, AI technologies have made notable progress in meteorology, offering faster and more energy-efficient predictions. Under typical atmospheric conditions, these systems can match or even surpass conventional numerical weather models in accuracy. However, the study demonstrates that this advantage does not extend to extreme scenarios, such as unprecedented heat waves, cold spells, or severe wind events.

The research team, led by Zhongwei Zhang from the Institute of Statistics at KIT, conducted a detailed analysis of how AI models perform when forecasting weather events that exceed historical records. Their results indicate that the high-resolution HRES model developed by the European Centre for Medium-Range Weather Forecasts consistently delivers more reliable predictions under such extreme conditions.

The study compared several prominent AI-based forecasting systems, including GraphCast, Pangu-Weather, and Fuxi, against the HRES reference model. While AI models performed well across general weather scenarios, they exhibited a clear tendency to underestimate both the intensity and frequency of record-breaking events. The discrepancy becomes more pronounced as the magnitude of the event surpasses previously observed data.

According to Zhang, this pattern reflects a structural limitation of AI systems, which rely heavily on historical datasets for training. Because extreme weather events often fall outside the range of past observations, these models struggle to extrapolate accurately beyond their training domain.

Sebastian Engelke, a professor at the University of Geneva, further explained that neural networks lack the ability to reliably predict conditions that have not been previously encountered. In contrast, physics-based models are grounded in fundamental atmospheric laws, allowing them to maintain predictive accuracy even when the climate system behaves in unprecedented ways. This distinction is particularly important as climate change increases the likelihood of extreme weather phenomena with significant societal and economic consequences.

The study also underscores the implications for early warning systems and disaster preparedness. Underestimating the severity of extreme weather could lead to delayed or insufficient warnings, increasing risks to public safety and infrastructure. The researchers caution against relying exclusively on AI for high-stakes forecasting applications and instead advocate for a combined approach that integrates both AI and traditional numerical models.

Looking ahead, the study outlines potential strategies to enhance AI forecasting capabilities. These include incorporating simulated extreme events into training datasets, applying advanced statistical methods tailored to extreme values, and developing hybrid models that merge physical principles with machine learning techniques.

The research involved contributions from multiple institutions, including ETH Zurich, the Helmholtz Centre for Environmental Research, Technische Universität Dresden, and the University of Geneva.

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/

Media Contacts:

Christian Könemann
Chief Press Officer
Phone: +49 721 608-41105
Fax: +49 721 608-43658
christian koenemann∂kit edu

Dr. Joachim Hoffmann
Press Officer
Phone: +49 721 608-41151
joachim hoffmann∂kit edu

SOURCE: Karlsruhe Institute of Technology