Disentangling reduced representations of teleconnections using variational autoencoders

Studying teleconnections using data-driven methods relies on identifying suitable representations of the relevant dynamical processes involved. Often, these representations are identified through a dimensionality reduction of the dynamical process itself, such as the Niño3.4 index to represent the El-Niño Southern Oscillation or the clustering of circulation regimes to represent states of the North Atlantic eddy-driven jet. The relationship between these representations can subsequently be assessed in a causal model. However, since these representations are identified independently of the teleconnection studied, they do not necessarily capture the dynamical processes relevant for explaining the relationship between the two phenomena. Here, we present a regularised dimensionality reduction approach using variational autoencoders, a deep generative machine learning method, to identify reduced representations of large-scale processes and their teleconnections jointly in a causal framework. Applying the approach to study regional dynamical drivers of precipitation extremes over Morocco at subseasonal lead times, we show that the method is able to identify a representation of the circulation over the North Atlantic, which disentangles the drivers of precipitation over Morocco while maintaining its subseasonal predictability and physical interpretability. Furthermore, we demonstrate the ability of the approach to disentangle large-scale teleconnections at longer lead times.