Accurate atomic scanning transmission electron microscopy analysis enabled by deep learning

Abstract

Currently, the machine learning (ML)-based scanning transmission electron microscopy (STEM) analysis is limited in the simulative stage, its application in experimental STEM is needed but challenging. Herein, we built up a universal model to analyze the vacancy defects and single atoms accurately and rapidly in experimental STEM images using a full convolution network. In our model, the unavoidable interference factors of noise, aberration, and carbon contamination were fully considered during the training, which were difficult to be considered in the past. Even toward the simultaneous identification of various vacancy types and low-contrast single atoms in the low-quality STEM images, our model showed rapid process speed (45 images per second) and high accuracy (> 95%). This work represents an improvement in experimental STEM image analysis by ML.