Why Traditional Weather Forecasting Still Outshines AI for Extreme Events: 10 Key Insights
AI weather forecasting fails at extreme events due to training data limitations; traditional models remain superior for record-breaking heat, wind, and cold.
Artificial intelligence is being hailed as the future of weather forecasting—faster, more precise, and increasingly capable. Yet, when it comes to the most dangerous and record-breaking weather events, AI still falls short. A recent study published in Science compared leading AI models like GraphCast and Pangu-Weather with traditional physics-based models, revealing a critical blind spot: AI struggles to predict extreme conditions. Here are 10 key insights into why traditional forecasting continues to outperform AI for the most extreme weather.
1. AI's Training Data Blind Spot
AI weather models are trained on decades of historical weather data. They learn patterns by analyzing past observations and then apply those patterns to predict future weather. However, extreme weather events—by definition—are rare and often outside the range of what has been recorded. As Sebastian Engelke, a statistics professor at the University of Geneva and co-author of the study, explains, “If we’re looking at extreme weather, and especially record-breaking events, then this has not been observed in the past. It’s really the lack of information in their training data that makes it almost impossible for them to forecast it.”
2. The Challenge of Record-Breaking Heat
Consider the heatwave that struck Siberia in early 2020, which led to devastating wildfires and melting permafrost. This event was nearly impossible without climate change—another study found global warming made it 600 times more likely. Yet AI predictions consistently underestimated the extreme temperatures. Traditional physics-based models, which simulate the atmosphere using mathematical equations, did a better job capturing the intensity because they are not limited by past data.
3. Underestimating Extreme Wind and Cold
AI’s shortcomings are not limited to heat. The study found that AI models also fare poorly when predicting extreme wind speeds and record-breaking cold snaps. For communities that depend on accurate warnings for hurricanes, blizzards, or windstorms, this is a serious concern. Traditional models, while far from perfect, provide more reliable forecasts for these high-impact events, giving emergency managers critical extra time to prepare.
4. The Fundamental Problem of Novel Events
At its core, AI is a pattern-matching machine: it reproduces what has happened before. When a weather event is unprecedented—whether due to climate change or natural variability—AI has no relevant pattern to draw upon. This fundamental limitation means that for the most dangerous weather, which is often the most novel, AI is inherently less trustworthy. Traditional forecasting, grounded in physics, can extrapolate beyond historical conditions by solving equations that govern the atmosphere.
5. Traditional Models Adapt Better
Physics-based models use complex mathematical representations of the Earth’s atmosphere, oceans, and land surfaces. They can incorporate new physical understanding and adjust to changing conditions more readily than AI. For example, they can simulate the effects of a warming climate on storm intensity even if no exact analog exists in history. This adaptability makes them especially valuable for extreme weather, where conditions can evolve rapidly and unpredictably.
6. AI Still Excels at Typical Weather
It’s not all bad news for AI. For everyday weather—such as a typical rain shower, a moderate temperature swing, or a standard wind pattern—AI can outperform traditional models in speed and accuracy. This is because these events are well-represented in training data. The study notes that AI excels for “extreme weather that isn’t wildly outside the range of past events.” So for your daily forecast, AI is a valuable tool.
7. Nvidia's Atlas and Storm Dennis Example
Nvidia recently released its AI forecasting model, Atlas. In a demonstration, Atlas successfully predicted Storm Dennis, a rapidly intensifying cyclone that impacted the UK, despite not being explicitly trained on that event. Mike Pritchard, Nvidia’s director of climate simulation research, noted that the model “captured realistically intense wind events and really intense cyclones that cause damage.” However, such success stories are still the exception rather than the rule for extreme events.
8. Hurricane Path Prediction: A Success
AI models have shown particular promise in predicting hurricane tracks—the path a storm will take. They can quickly process vast amounts of data and generate accurate trajectories, often faster than traditional models. This has led to their adoption by agencies like the National Hurricane Center and companies like the Weather Company. Yet even here, intensity forecasts (how strong the storm will be) remain a challenge for AI, especially for rapid intensification.
9. AI Models Are Improving but Not There Yet
The study examined AI models as they existed about a year ago. Since then, improvements have been made, including the addition of probabilistic models that generate multiple possible outcomes to better capture uncertainty. However, the fundamental reliance on historical training data remains. As Engelke points out, “the basic problem still exists because they are still based on training data from the past.” So while AI is improving, it has not yet closed the gap for extreme events.
10. The Future: Hybrid Approaches?
Given the complementary strengths of AI and traditional forecasting, many experts advocate for hybrid systems. AI can handle routine forecasts quickly, while physics-based models tackle extreme and novel events. Weather agencies already use both in parallel, comparing outputs to produce the best forecast. As AI continues to evolve—especially with techniques that incorporate physical constraints—this blend may offer the most reliable path forward for protecting lives and property from extreme weather.
In conclusion, while AI has made remarkable strides in weather forecasting, its inability to predict record-breaking extreme events is a significant limitation that cannot be ignored. Traditional physics-based models remain the gold standard when the stakes are highest. For now, the smartest approach is to use AI for what it does best—routine forecasting—and rely on time-tested methods for the extremes that matter most.