A recent study confirms that classical numerical models remain superior to artificial intelligence models for predicting extreme weather phenomena, a crucial issue for safety and climate adaptation.
The Announcement
An analysis published by Carbon Brief reveals that, despite the spectacular advances of artificial intelligence (AI) models, traditional numerical models continue to provide better forecasts for extreme weather events. These results highlight the current limitations of AI when faced with the complexity of high-intensity atmospheric phenomena.
According to this study, classical models, such as those developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), remain the benchmark for anticipating climate records and high-impact weather situations. AI, while promising, has not yet matched their accuracy in these critical cases.
What We Know
Traditional models rely on the fine resolution of the dynamic equations of the atmosphere, integrating satellite data and ground observations to rigorously simulate weather developments. By comparison, predictive models based on machine learning, such as GraphCast, Pangu-Weather, or FourCastNet, use neural networks to learn from historical patterns and generate forecasts.
Researchers emphasize that while AI excels in speed and generating global scenarios, it struggles to reproduce the complexity of physical interactions during extreme events. Classical numerical models retain a better ability to model these phenomena thanks to their scientific foundation and the thorough integration of atmospheric processes.
This superiority does not call into question the interest of AI in meteorology but highlights the need for complementary development between traditional approaches and artificial intelligence.
Why It Matters
Accurate forecasts of extreme events — hurricanes, heatwaves, violent storms — are essential to protect populations, anticipate risks, and adapt infrastructures. An error or delay in forecasting can lead to dramatic consequences in terms of civil and economic safety.
In a context of climate change where the frequency and intensity of extreme phenomena are increasing, having reliable tools is a priority. This study reminds us that, for now, numerical physical models remain indispensable to guarantee robust forecasts during these major meteorological crises.
The Community's Reaction
The scientific and meteorological community welcomes these results as a call for caution and cooperation. Rather than considering AI as a replacement for classical methods, experts advocate for synergy between physical modeling and machine learning.
Specialists emphasize that AI can greatly improve data processing speed and global coverage, but the scientific rigor of traditional models remains crucial for the reliability of extreme alerts.
The Technical Challenges of AI in Extreme Meteorology
One of the main obstacles AI faces in forecasting extreme events is the scarcity of representative data. Extreme phenomena, by definition, occur infrequently, which limits the volume of historical data available for model learning. This insufficiency complicates neural networks' ability to generalize correctly and predict unprecedented or exceptionally intense situations.
Moreover, the physical complexity of atmospheric interactions — such as nonlinear feedbacks between the ocean, atmosphere, and land surface — requires precise modeling of fundamental physical laws. Traditional models explicitly integrate these principles, giving them an advantage in forecast fidelity during extreme conditions. In contrast, AI tends to operate as a black box, which can limit understanding and interpretation of predicted phenomena.
Finally, the spatial and temporal resolution of AI models still needs improvement to capture the fine details of violent weather events. While classical models can reach fine grids and simulate microphysical processes, neural networks must still advance to match this level of detail without sacrificing execution speed.
Perspectives on Integrating AI and Traditional Modeling
Faced with these limitations, researchers are exploring hybrid strategies combining the strengths of AI and physical models. For example, AI can be used to accelerate certain calculation steps, such as data assimilation or bias correction, while retaining the physical bases of the numerical model. This approach would combine speed and reliability, two essential criteria for managing risks related to extreme phenomena.
Furthermore, AI can contribute to post-processing results by refining anomaly detection or generating probabilistic scenarios to better apprehend uncertainty. This collaborative work between disciplines paves the way for more efficient and adaptive meteorological tools in the face of challenges posed by climate change.
Ongoing projects, supported by international institutions, aim to develop these hybrid models by 2027-2028, with strong ambitions to improve the accuracy and speed of extreme forecasts worldwide.
In Summary
Despite the rapid progress of artificial intelligence, traditional weather models remain to date the most reliable for forecasting extreme events. Their rigorous scientific foundation and ability to finely model physical processes give them a major advantage. AI, for its part, offers new opportunities in terms of speed and global coverage but still must overcome significant challenges related to the complexity of phenomena and data scarcity.
The future path likely lies in the complementarity and integration of these two approaches to improve climate risk management. By combining the rigor of numerical models and the processing power of AI, it will be possible to design more effective meteorological tools, essential in the face of the intensification of extreme phenomena due to global warming.
