Nature has published a story highlighting research by the DSI’s AI for Climate (AICE) Research Initiative on using artificial intelligence to forecast extreme weather events.

The article focuses on a series of studies by AICE researchers including Pedram Hassanzadeh (Faculty Director, AICE; Assoc. Prof, Geophysical Sciences), Tiffany Shaw (Faculty Director, AICE; Prof, Geophysical Sciences), research scientist Y. Qiang Sun, Schmidt AI in Science Fellow Alexander Wikner, and collaborators including Dorian Abbot (UChicago), Jonathan Weare (NYU), and others from the École Normale Supérieure, Paris.

The team is tackling the challenge of predicting “gray swan” events: rare, extreme weather occurrences that may happen only once every 1,000 years, but cause devastating impacts when they do. With climate change making such events more frequent, knowing how to forecast them is increasingly urgent. But when AI weather models typically train on just 40 years of historical data, how can they predict something they’ve never seen?

In their analyses, the researchers found that current AI models struggle to extrapolate to unprecedented extremes on their own. But combining AI with physics-based climate models offers a promising path forward: in early tests, this hybrid approach achieved the accuracy of traditional methods at higher speeds.

“Maybe next year there’s another event that never happened before in history,” Sun said. “We would like to understand how to be better prepared.”

The team will present their findings at the American Geophysical Union annual meeting next week.

Read the original story here.

For a deeper technical discussion, see the researchers’ invited article in SIAM News: “Forecasting the Unseen: AI Weather Models and Gray Swan Extreme Events.”