Machine learning based decadal time scale AMOC predictions

Abstract

Both ocean internal variability and external forcing are considered to be important for decadal predictions for the Atlantic Meridional Overturning Circulation (AMOC). Here we employ a Convolutional Neural Network (CNN) based model with surface/subsurface temperature and radiative forcing as input to predict the AMOC strength. We train, validate and test the CNN on simulations of one Earth System Model (ESM), achieving skillful AMOC predictions for decadal lead times. Generalization of the CNN including transfer learning achieves skill in AMOC predictions of another ESM. Using ORAS5 reanalysis and RAPID data as input, future (2024-2033) CNN predictions have comparable skill compared to those from ESMs, but at a much lower computational cost. An Explainable AI analysis of the CNN highlights the importance of radiative forcing and zonal subsurface temperature gradients as key physics captured by the CNN. Our results may help improve the skill of decadal AMOC predictions, important for future climate change.