Extreme weather events—floods, heatwaves, and storms—cause billions in damages annually, making accurate forecasting a life-saving priority. For decades, physics-based numerical weather prediction has been the gold standard. But AI weather models have recently surged, promising faster, cheaper forecasts. However, a new study in Science Advances delivers a crucial reality check: when it comes to record-breaking extremes, traditional models still outperform AI. Here are eight key insights from that research, each highlighting why replacing old-school forecasting with AI too quickly could be a costly mistake.
1. Traditional Models Still Rule for Record-Breaking Extremes
Despite AI's impressive gains in routine weather prediction, the study confirms that physics-based models remain superior for forecasting record-breaking events. The researchers tested both systems on thousands of extreme hot, cold, and windy events recorded in 2018 and 2020. Traditional models consistently delivered more accurate simulations of these rare, high-impact phenomena. This isn't a small gap—AI systematically underestimated both the frequency and intensity of such extremes. For communities bracing against unprecedented heat or storms, relying solely on AI could lead to underpreparedness and increased risk.
2. The New Study's Surprising Findings
Published in Science Advances, the research directly compared AI and traditional models head-to-head for extreme event prediction. The results were stark: AI models failed to capture the full magnitude of rare, record-breaking weather. While AI excels at patterns found in its training data, extreme events are, by definition, outliers. The study's lead author, Prof. Sebastian Engelke from the University of Geneva, calls the finding a “warning shot” against prematurely replacing physics-based models. The implication is clear—even as AI advances, it cannot yet handle the most dangerous weather.
3. Why AI Underestimates Record-Breaking Events
The root cause lies in how AI learns. These models are trained on historical weather data, teaching them to recognize common patterns. But record-breaking events are rare and often fall outside the range of training data. Engelke explains that AI models “depend strongly on the training data” and are “relatively constrained to the range of this dataset.” This means they struggle to simulate brand-new extremes—a fundamental limitation. Traditional physics-based models, which rely on mathematical equations grounded in natural laws, can extrapolate beyond historical records because they simulate physical processes, not just past patterns.
4. How Physics-Based Models Have Long Dominated Forecasting
Numerical weather prediction models have been the backbone of forecasting for decades. They use complex equations that replicate atmospheric and oceanic processes—equations rooted in fundamental physics. These models don't need to have seen an event before; they compute how the atmosphere should behave based on current conditions and physical laws. This makes them inherently capable of predicting unprecedented scenarios. While they require significant computational power, their accuracy for extreme events has been repeatedly validated, making them indispensable for early warning systems worldwide.
5. The Rise of AI Weather Models
In recent years, AI has stormed into weather forecasting. Instead of physics, these models use statistical techniques, learning from massive historical datasets. They recognize patterns and make predictions by finding similar past situations. AI models have outperformed traditional ones in many routine forecasts—especially for temperature and precipitation over short timeframes. They also use far less computing power, making them cheaper and faster. This has led some to champion AI as the future of meteorology, but the new study highlights a critical blind spot.
6. The Advantages That Make AI Attractive
AI weather models bring real benefits. Their lower computational demands allow forecasts to run on ordinary computers rather than supercomputers. They can deliver results in minutes instead of hours. Many AI models also show superior performance on average conditions and common weather patterns. For everyday forecasting—like next week's temperatures or rain chances—AI often beats physics-based models. This practical success has spurred investment and optimism. Yet the trade-off becomes apparent when rare, extreme events push the models beyond their training data's limits.
7. The Critical Drawback: Training Data Limitations
The very strength of AI—learning from past data—becomes its weakness with extremes. If a heatwave is historically unprecedented, the AI has no similar pattern to reference. It will tend to underestimate because its training data doesn't include such high values. This is not a simple fix; adding more data doesn't solve the problem because record-breaking events are inherently rare. Traditional models, by contrast, simulate the physics of heat build-up regardless of whether such a combination has occurred before. This fundamental difference means AI and physics-based models should be seen as complementary, not replacements.
8. A Warning Against Replacing Traditional Models Too Quickly
Prof. Engelke's “warning shot” is clear: do not discard physics-based models prematurely. While AI can enhance forecasting—especially for routine conditions—it is not yet ready to handle the extremes that cause the most damage. The study urges a hybrid approach, where traditional models continue to lead for severe weather, and AI supplements them for faster, cheaper routine forecasts. Governments and agencies investing in early warning systems should maintain the proven physics-based tools while thoughtfully integrating AI's strengths. Rushing to replace traditional models could leave society dangerously exposed to the next record-breaking storm or heatwave.
In conclusion, the new research underscores a vital lesson: AI is a powerful tool, not a magic bullet. Traditional physics-based models remain the gold standard for extreme weather forecasting, precisely because they understand the atmosphere's fundamental rules. The future likely lies in combining both approaches—using AI for speed and efficiency, while keeping physics-based models as the backbone for the most dangerous events. For now, any move to fully replace traditional models with AI should be met with caution, especially when lives and livelihoods are at stake.