Item

Abstract

Machine Learning is fast becoming a crucial tool to process and investigate the big data of Ocean General Circulation Models (OGCM), and together with explanation methods it is possible to look into the decisions of learning methods, and thus understand dynamics. Even though monitoring arrays like RAPID or OSNAP are large international efforts, their limited sampling resolution still challenges the deciphering of AMOC variability into local vs. remote causes. Patterns that are correlated with these remote causes are of great interest for the community and are key for understanding the AMOC outside of monitoring arrays.To extract possible patterns and causes for the AMOC variability which can be later tested against single observations, we utilize OGCM data from the eddy-rich ocean model VIKING20X to train a Neural Network (NN). Its output is subsampled to mimic sparse observations (e.g., ARGO profiles), and explored to predict reconstructions of the complete fields and the AMOC strength of the model ‘truth’. By making the trained NNs interpretable with explainable artificial intelligence, we extract maps of interest that indicate which patterns are important for ocean dynamics and AMOC variability. To avoid a direct comparison between the model interpretations and the real world observations we aim for highly generalizable knowledge extracted from our NN.In conclusion, this study identifies the usefulness of applying NNs to aid the observational sampling strategy and the interpretation for AMOC monitoring.