Conventional weather models predict extreme events better than AI models
AI models can now predict the weather very precisely. However, they have weaknesses in extreme events with record values - for obvious reasons.
Temperature anomalies during the record heatwave in Siberia in 2020
(Image: Zhongwei Zhang, KIT)
Artificial intelligence (AI) has revolutionized weather forecasting. With significantly fewer resources than conventional methods, AI algorithms can predict with high accuracy – with one exception, says a team from the Karlsruhe Institute of Technology (KIT) and the University of Geneva.
AI models like WeatherNext 2 from Google's AI division Deepmind are now at least on par with, if not superior to, physics-based numerical models in terms of accuracy. For example, WeatherNext 2 can provide forecasts for two weeks regarding temperature, air pressure, and wind with an accuracy of one hour – previous systems could only do this for two days. WeatherNext 2 requires less computing power.
However, when it comes to extreme weather events with record temperatures, wind speeds, or precipitation, AI weather models have proven to be less powerful, as the KIT team led by Zhongwei Zhang has found out. Here, the physics-based high-resolution model HRES from the European Centre for Medium-Range Weather Forecasts (ECMWF) is still superior.
The team compared the results of several AI models, including GraphCast, Pangu Weather, and Fuxi, with those of HRES. The AI models performed well in the overall assessment of all weather situations.
Errors in predicting record values
However, their forecasts were flawed for record events. This applied particularly to the frequency of extreme events and their intensity. “Our analyses show that AI models generally underestimate the intensity of heat, cold, and wind records,” said project leader Zhang. “The more a record exceeds previous extreme values, the greater the underestimation.”
The researchers say this is inherent: AI models learn from historical data. Their strength lies in predicting patterns that resemble previously experienced situations. However, record events, which are occurring with increasing frequency due to climate change, are outside previous experience.
“Neural networks have difficulty reliably extrapolating beyond their training range – meaning making predictions beyond previously observed values,” said Sebastian Engelke from the University of Geneva. “Physical models like HRES, on the other hand, are based on fundamental laws of physics. This ensures that their predictions remain reliable even when the atmosphere transitions into states that have not yet been observed.”
Videos by heise
The results are relevant for early warning systems, the researchers emphasize: If extreme events are underestimated, it may lead to warnings being issued too late or not at all. Therefore, AI weather models cannot currently replace numerical forecasts, the team writes in the journal Science Advances. “For high-risk applications, one should not rely solely on AI,” Zhang concluded. Both approaches should be used in parallel.
(wpl)
