Real-space analysis of scanning tunneling microscopy topography datasets using sparse modeling approach

A sparse modeling approach is proposed for analyzing scanning tunneling microscopy topography data, which contain numerous peaks originating from the electron density of surface atoms and=or impurities. The method, based on the relevance vector machine with L₁ regularization and k-means clustering, enables separation of the peaks and peak center positioning with accuracy beyond the resolution of the measurement grid. The validity and efficiency of the proposed method are demonstrated using synthetic data in comparison with the conventional least-squares method. An application of the proposed method to experimental data of a metallic oxide thin-film clearly indicates the existence of defects and corresponding local lattice distortions.