Using $k$-means to sort spectra: electronic order mapping from scanning tunneling spectroscopy measurements
Hyperspectral imaging techniques have a unique ability to probe the inhomogeneity of material properties whether driven by compositional variation or other forms of phase segregation. In the doped cuprates, iridates, and related materials, scanning tunneling microscopy/spectroscopy (STM/STS) measurements have found the emergence of pseudogap 'puddles' from the macroscopically Mott insulating phase with increased doping. However, categorizing this hyperspectral data by electronic order is not trivial, and has often been done with ad hoc methods. In this paper we demonstrate the utility of $k$-means, a simple and easy-to-use unsupervised clustering method, as a tool for classifying heterogeneous scanning tunneling spectroscopy data by electronic order for Rh-doped Sr$_2$IrO$_{4}$, a cuprate-like material. Applied to STM data acquired within the Mott phase, $k$-means successfully identified areas of Mott order and of pseudogap order. The unsupervised nature of $k$-means limits avenues for bias, and provides clustered spectral shapes without a priori knowledge of the physics. Additionally, we demonstrate successful use of $k$-means as a preprocessing tool to constrain phenomenological function fitting. Clustering the data allows us to reduce the fitting parameter space, limiting over-fitting. We suggest $k$-means as a fast, simple model for processing hyperspectral data on materials of mixed electronic order.