A recent study conducted by researchers at the University of Geneva and the Karlsruhe Institute of Technology compares AI-based weather forecasting with traditional physics-based models.
The main message is that artificial intelligence can forecast routine weather well, but it consistently underestimates the intensity and frequency of extreme events such as heatwaves, heavy rainfall, and violent storms, even up to ten days ahead.
This has immediate implications for early warning and disaster-management efforts in a world where climate change is making extremes more common.
The research, published in Science Advances, highlights a critical gap between cutting-edge AI approaches and physics-based numerical weather prediction that decision-makers cannot afford to ignore.
Key findings of the study
From a long-standing perspective in weather prediction, the study confirms that AI systems excel at identifying familiar, recurring patterns in historical data.
However, when it comes to rare or unprecedented extremes, AI models show a troubling blind spot.
The core result is that AI tends to miss both the intensity and the frequency of record-breaking events, particularly when forecasting more distant timescales of up to ten days.
AI excels at routine forecasts
In everyday weather prediction tasks, AI-based models can efficiently process vast data streams and deliver accurate short-term forecasts.
This capability aligns with their strength in pattern recognition and speed, offering potential cost and workflow advantages for standard weather prediction tasks.
AI struggles with extremes
When the forecast demands rise above routine conditions, AI encounters a fundamental limitation.
Trained on historical data from 1979 to 2017, AI systems implicitly cap how extreme outcomes can appear, effectively underestimating both the severity and the rarity of novel events that climate change is increasingly producing.
The consequence is a lower reliability for early warning and disaster-management decisions, where missing an extreme can have life-or-death consequences.
Why physics-based models remain essential
Physics-based or numerical weather prediction (NWP) models are built on the physical laws governing atmospheric dynamics.
Despite higher computational costs, these models can reproduce weather states that have never occurred in the historical record, making them better suited for forecasting unprecedented extremes.
The study reinforces the view that a strong theoretical foundation enables physics-based methods to generalize beyond past observations—a robustness AI currently lacks for extreme events.
Strengths of physics-based models
By solving the fundamental equations of motion, thermodynamics, and mass conservation, physics-based models can represent dynamical processes with a level of fidelity that AI struggles to match.
This makes them more trustworthy for high-stakes forecasts and for scenarios involving extreme or novel weather patterns, even if they demand more computational resources and expert calibration.
Implications for climate resilience and policy
The urgency of improving forecasts for extreme events grows as climate change elevates the frequency and intensity of record weather.
The current gap between AI and physics-based forecasting is not just a technical curiosity; it is a real-world risk for emergency planning, insurance, infrastructure design, and disaster response.
While AI could streamline routine forecasting and reduce costs, those gains cannot come at the expense of reliably predicting dangerous conditions.
Practical recommendations for decision-makers
- Invest in hybrid forecasting systems that combine AI pattern recognition with physics-based constraints to improve extreme-event predictions.
- Maintain strong uncertainty quantification and clearly communicate confidence levels to authorities.
- Continue using physics-based models as the backbone for high-stakes forecasts, especially near expected extremes.
- Stress-test AI models against synthetic extreme scenarios to evaluate limits and avoid overreliance on historical patterns.
- Ensure ongoing validation with the most recent data and monitor shifts in climate regimes that may outpace historical experience.
What the industry should watch for next
Experts advocate for significant improvements in AI models before they can operate independently in critical systems.
The era of “old guard” physics-based prediction remains essential for high-stakes forecasting, even as AI offers promising gains in efficiency and automation.
Closing the gap will require interdisciplinary collaboration, robust testing against extremes, and a commitment to transparent risk assessment.
Conclusion: toward a safer, more reliable forecast future
As weather extremes become more frequent in a warming world, forecasting tools must deliver both accuracy and reliability across routine and extreme conditions.
The study from Geneva and Karlsruhe serves as a cautionary benchmark: AI can support the prediction enterprise, but it cannot replace the foundational strength of physics-based models in critical, life-saving applications.
Here is the source article for this story: AI can do almost everything except predicting extreme weather events: Why?