What Type of Weather Model is Better at Predicting Extreme Weather Events? A New Study Has the Answer
AI is making its way into weather forecasting. AI models such as GraphCast, Pangu-Weather and Fuxi are already better than traditional physics-based climate models for daily weather prediction, though they still aren’t perfect.
A new study published in Science Advances stated that AI often fails to predict record-breaking extreme weather events. Extreme weather events, like record heat waves and windstorms, are becoming more frequent with the changing climate. Having accurate warnings is essential to help protect lives, property and infrastructure. The unprecedented nature of these kinds of events is an issue for AI.
New Research
Scientists compared leading AI models against HRES (High Resolution Forecast), which is one of the world’s leading physics-based weather prediction systems. They built a large database of record-breaking heat, cold and wind events from 2018 and 2020. Researchers cross-referenced the forecasts that HRES and the AI models had made for those years to see which one got the closest to the real outcome.
AI models are often more accurate for everyday weather forecasts as well as much. Faster than HRES. However, HRES very much outperformed AI in record-breaking events. For record-breaking heatwaves, AI models consistently predicted temperatures to be much lower than what was observed. The more a record was broken, the less accurate AI was.
HRES was performing better in these situations due to its foundation with the laws of physics. The laws of physics never change. Physics-based models have the ability to better simulate scenarios that the world has not experienced. AI models faced with events not included in their training data were trying to compensate with historical averages instead.
What the Scientists Are Saying
“Our findings underscore the current limitations of AI weather models in extrapolating beyond their training domain and in forecasting the potentially most impactful record-breaking weather events,” explains the research team.
The researchers warn against becoming fully reliant on AI for important work like this given the fact that extreme events will become more frequent. They suggest using a hybrid approach to combine the speed of AI with the strong foundation of the laws of physics. “Further rigorous verification and model development is needed before these models can be solely relied upon for high-stakes applications such as early warning systems and disaster management,” says the research team.
“Zhongwei Zhang et al, Physics-based models outperform AI weather forecasts of record-breaking extremes, Science Advances (2026). DOI: 10.1126/sciadv.aec1433”