AI emulator highlights underestimated risk of AMOC collapse in current climate models

The pivotal role in regulating the global climate system and the potential for irreversible collapse underscore the critical importance of the Atlantic Meridional Overturning Circulation (AMOC) and its stability. To address the inherent nonlinearity and stochastic nature of the AMOC, we develop a Convolutional Neural Network (CNN) model to project AMOC evolution using atmospheric and oceanic climate model inputs. Our CNN model successfully captures the stochastic AMOC bifurcation present in large-ensemble climate model simulations. Using explainable AI, we find that the salinity structure enables the CNN to predict future AMOC trajectories, suggesting that the salt-advection feedback amplifies subtle perturbations. Current climate models systematically misrepresent this salinity structure. We show that correcting these biases shifts the climate model projections towards a collapse-prone regime, implying that the AMOC’s stability in current climate models is likely overestimated. Our findings suggest that the risk of AMOC collapse cannot be ruled out simply based on model projections, calling for more thorough investigations of AMOC stability with focus on potential stability biases in climate models.