A Physics-Informed Gaussian Mixture Neural Network to Extract Atomic Signals from Scanning Tunneling Microscope Images

Scanning Tunneling Microscopes (STMs) have led to many scientific breakthroughs in nanoscience. The STMs cannot achieve their full imaging potential due to signal overlaps and interference. This project seeks to treat the signal interference using a Gaussian Mixture Neural Network (GMNN) with physics-based constraints incorporated. This method is able to localize the individual atomic signals within an STM image, providing a quantitative analysis of nano-material properties. The GMNN is applied to a case study of Graphene Nano-Ribbon (GNR). The underlying atomic structures of GNR are successfully extracted from blurry STM images. Additionally, the extracted atomic signals of GNR from STM images provide a unique way to measure the pairwise atom distances, which enables us to gain important insight into the irregularity of experimental GNR structures. This will allow us to study the correlation between GNR structure irregularity and its superconductivity and superfluidity in the future.