Caltech AI learns to spot sudden shifts ahead of time in simulated climate and airflow
Some systems change slowly and then flip suddenly, a moment scientists call a tipping point, and these shifts are hard to forecast. A Caltech-led team working with the University of Colorado Boulder and NVIDIA trained a recurrent neural operator, a neural network with a built-in memory of how a system’s physics is changing. Caltech reports that when the network’s predictions stop following the physics it learned, that signals a tipping point is near. In simulated data of low ocean clouds disappearing, it looked more than two years ahead and predicted the timing within about 11 days. It also anticipated a sudden drop in lift in simulated airflow over a wing. The study appears in PNAS.