Estimating the distance to the AMOC tipping point using convolutional neural networks

The Atlantic Meridional Overturning Circulation (AMOC) is a critical component of the Earth system and one of the most prominent tipping element. In a warming climate, the AMOC is at risk of collapse due to increased freshwater input in the North Atlantic. Such an extreme event could lead to severe consequences for the global climate, with strong socio-economics impacts. Such a tipping event has been demonstrated to occur in conceptual, intermediate complexity and, recently, in the Community Earth System Model (CESM). Therefore, Reliable early warning signals are required for detecting whether the AMOC is approaching a tipping point. To estimate the distance of the AMOC to tipping, we propose a novel methodology, based on a Convolutional Neural Network (CNN) which uses sea surface salinity and temperature across the Atlantic as input. First, we validate our approach within the model of intermediate complexity Climber-X, demonstrating its ability to generalize to different forcing rates and in the presence of noise. We also explore the use of alternative climate variables such as the full-depth salinity profile at 35°S. Second, we assess the generalization capability of our methodology to a model of higher complexity. To this end, we use the CNN trained on Climber-X and successfully apply it to the AMOC collapse recently simulated in the CESM model. To demonstrate the physical consistency of the CNN model and increase its interpretability, we identify the most relevant regions to estimate the distance of the AMOC to tipping via the Layer-wise Relevance propagation technique.