Atomic-scale analysis of disorder by similarity learning from tunneling spectroscopy

Abstract

Rapid proliferation of hyperspectral imaging in scanning probe microscopies creates unique opportunities to systematically capture and categorize higher dimensional datasets, toward new insights into electronic, mechanical and chemical properties of materials with nano- and atomic-scale resolution. Here we demonstrate similarity learning for tunneling spectroscopy acquired on superconducting material (FeSe) with sparse density of imperfections (Fe vacancies). Popular methods for unsupervised learning and discrete representation of the data in terms of clusters of characteristic behaviors were found to produce inconsistencies with respect to capturing the location and tunneling characteristics of the vacancy sites. To this end, we applied a more general, non-linear similarity learning. This approach was found to outperform several widely used methods for dimensionality reduction and produce a clear differentiation of the type of tunneling spectra. In particular, significant spectral weight transfer likely associated with the electronic reconstruction by the vacancy sites, is systematically captured, as is the spatial extent of the vacancy region. Given that a great variety of electronic materials will exhibit similarly smooth variation of the spectral responses due to random or engineered inhomogeneities in their structure, we believe our approach will be useful for systematic analysis of hyperspectral imaging with minimal prior knowledge, as well as prospective comparison of experimental measurements to theoretical calculations with explicit consideration of disorder.