Unsupervised Learning of Spatially-Resolved ARPES Spectra for Epitaxially Grown Graphene via Non-negative Matrix Factorization

Abstract

This study proposed an unsupervised machine-learning approach for analyzing spatially-resolved ARPES. A combination of non-negative matrix factorization and k-means clustering was applied to spatially-resolved ARPES spectra of the graphene epitaxially grown on a SiC substrate. The Dirac cones of graphene were decomposed and reproduced fairly well using non-negative matrix factorization (NMF). The base and activation matrices obtained from the NMF results reflected the detailed spectral features derived from the number of graphene layers and growth directions. The spatial distribution of graphene thickness on the substrate was clearly visualized by the clustering using the activation matrices acquired via NMF. Integration with k-means clustering enables clear visualization of spatial variations. Our method efficiently handles large datasets, extracting spectral features without manual inspection. It offers broad applicability beyond graphene studies to analyze ARPES spectra in various materials.