Investigating Decadal Variability in the North Atlantic Using a New Time-evolving Self-organizing Maps Method

Abstract:

The evolution of internal variability in the North Atlantic on decadal time scales has a critical impact on global and regional climate. Traditional methods of identifying variability are limited by their inability to represent the full spatio-temporal variability of the system, which is important for understanding the coupled dynamics in the North Atlantic. This study develops a novel method of representing the spatio-temporal evolution of internal variability using a time-evolving self-organizing maps (SOMs), a machine learning method that classifies high dimensional datasets into a specified number of clusters but in this case through providing a time series of input data. In this study, we use a long Community Earth System Model 1850 pre-industrial simulation to identify the evolution of sea surface temperature (SST) variability. The SOM is repeatedly trained using 10-year time series of winter SST in the North Atlantic sampled from 1500 years of the model simulation. The resulting SOM nodes identify the dominant patterns of 10-year continuous spatio-temporal evolutions of winter SST. These results show spatial shifts between North Atlantic Oscillation-like tripole and Atlantic Multidecadal Variability-like monopolar SST anomalies. In addition to these anomalies in the extra-tropical and tropical North Atlantic, this study also highlights the role of Greenland-Iceland-Norwegian Seas in the internal variability in the North Atlantic. The coupled atmosphere-ocean-sea ice mechanisms involved in producing these SST evolutions are then investigated through compositing multiple variables associated with deep water formation, air-sea heat flux, sea ice melt/growth, and ocean heat transport. The identification of underlying time-evolving coupled dynamics contributes to the understanding of SST evolution, bringing new insights that can benefit future studies on decadal predictability and actual decadal predictions.