AI climate model forecasts tipping points with neural operator

Researchers led by Caltech published in the Proceedings of the National Academy of Sciences a study showing that a recurrent neural operator can forecast critical tipping points in complex systems, including climate and aerodynamic simulations. The AI model predicts the loss of low ocean clouds from rising carbon dioxide and abrupt stalls in airflow over a tilting airplane wing. Tipping points describe thresholds at which a system becomes nonstationary and shifts into a new state, often irreversibly, and previous AI weather models struggled because they relied on stationary assumptions. The approach, developed with researchers from Caltech, the University of Colorado Boulder and NVIDIA, builds on neural operators by adding a memory mechanism that tracks how a system's governing physics change over time. When the AI's predictions no longer conform to previously learned physical rules, the method flags an approaching threshold. In simulated cloud data, the recurrent neural operator forecast the expected cloud-loss tipping point more than two years in advance, with a timing error of about 11 days. In wing airflow tests, it predicted the moment when lift drops sharply, a signal that could help anticipate aerodynamic instabilities in aircraft design. Principal author Anima Anandkumar, Caltech's Bren professor of computing and mathematical sciences, said the work establishes a foundation for characterizing tipping points and estimating when they may occur in nonstationary systems. The method matters because abrupt changes in climate, ecology, epidemiology and aerospace systems can become catastrophic once a point of return is crossed. The study was also cited in a recent United Nations report on the importance of tipping points in climate science.