Item

Abstract

Despite the importance of the Atlantic Meridional Overturning Circulation (AMOC) to the climate, direct measurements are sparse. Before the onset of the current century, many reconstructions of the AMOC rely on linear relationships to the more readily observed surface properties of the Atlantic rather than the temporal sparsely observed subsurface. Sustained records of the AMOC by monitoring arrays are available for almost two decades. However, due to the large number of experiments performed with Earth System Models (ESM), we have large datasets at our disposal that allow us to apply machine learning methods to investigate how robust short sustained observational records can be reconstructed and determine the relative importance of surface and subsurface information for the reconstruction.We train different linear and non-linear machine learning models to infer the AMOC from sea-water properties at different depths, using large ensemble simulations with the Max Planck Institute ESM. During training, we retain consecutive periods of our data for the model evaluation resembling the timescale of our observational record. Subsequently, we investigate the validity of these reconstructions with explainable Machine Learning techniques that map the relevance of the reconstruction back to the model input. We identify regions in the surface and subsurface ocean which appear to be relevant for the reconstruction of the AMOC and compare the quality of the reconstructions using either only surface or only subsurface information. We establish the quality of AMOC reconstructions based on only linear relationships and how robust these reconstructions are to different retained periods.