Artificial intelligence has become an important tool for weather forecasting. AI models often outperform conventional forecast models, providing more detail and the ability to look farther into the future.
AI forecasting systems make predictions by identifying patterns in historical weather data. Conventional weather models, on the other hand, are based on atmospheric physics.
The historic blizzard that blanketed the northeastern US in February showed that AI forecasting still has some shortcomings. The Global Forecast System, which is the conventional U.S. weather model, warned that there would be heavy snowfall several days before the storm arrived. AI models were less certain. As it turned out, the blizzard dumped 20 inches of snow on New York’s Central Park, which made it the ninth biggest snowstorm ever to hit Manhattan.
The storm was what people call a gray swan. Gray swans are high-impact, potentially catastrophic events that are foreseeable yet assumed to be unlikely. This is distinguished from common risks as well as from black swans, which are completely unexpected events.
Research by the University of Chicago looked at the ability of AI models to deal with extreme weather events. It showed that AI models might fail to predict storms that have little precedent in the weather record. This problem could become more acute in the years ahead as the changing climate fuels ever more extreme weather.
To help AI models better predict grey swan weather events, it may be that physics needs to be built into the models. If the models can really learn atmospheric dynamics, they may be less likely to be surprised by extreme weather events.
