On using self-organizing maps in analyses of circulation types and circulation modes

Self-organizing maps (SOMs) represent an increasingly popular method in climatology. They are able to reduce very complex atmospheric circulation data into a two-dimensional array of patterns, or provide a topologically-ordered nonlinear classification of high-dimensional data. As such, their output combines weather/circulation types with exploratory projection / dimensionality reduction, that is, two approaches to feature extraction that are usually conceptualized and interpreted in very different ways and studied by sets of rather distinct methods. As a consequence to this complexity of SOMs, they have been interpreted not only as weather/circulation types, but, alternatively and less intuitively, the array has also been suggested to represent a continuum of teleconnections. The distinct characteristics of SOMs, in addition to the absence of a widely accepted framework for defining teleconnections and a standard set of terminologies, can potentially cause confusion and misinterpretations when comparing works of authors with different perspectives. This is particularly true when reading the conclusions of studies without a thorough understanding of the context. Our contribution briefly introduces SOMs and their relation to circulation types, circulation regimes, modes of variability, and teleconnections. We use synthetic datasets generated from idealized modes of variability and from main modes of Euro-Atlantic atmospheric circulation variability extracted from daily SLP fields by a rotated principal component analysis to demonstrate how SOM arrays capture the spatial patterns of modes as well as changes in the strength and phase of modes. Additionally, we explore the possibility of using linear feature-extraction tools such as PCA to train high-quality SOMs more efficiently.