Atom/defect classification in Scanning Transmission Electron Microscopy (STEM) Images using Convolutional Neural Networks

Abstract

Two-dimensional (2D) materials have shown great capabilities in various research fields such as physics, chemistry, and engineering because of their low dimensionality and versatility of behavior as compared to their bulk counter parts. Parameters such as twist angles between layers, defect concentrations, and strain can drastically change the properties of the 2D material. To study the interactions inside the 2D material at an atomic scale, the use of electron microscopy techniques such as scanning transmission electron microscopy (STEM) is prevalent in their characterization. However, a bottleneck is hit when it comes to interpreting the STEM image data because of the extent of manual labor and subjective decisions required to segment the atomic columns. Here, a novel approach is proposed for atom/defect classification in STEM images of twisted 2D bilayer materials using deep learning-based image segmentation model. Since it is not feasible to generate a training dataset experimentally, a Python library was developed that leverages the Atomic Simulation Environment (ASE) and abTEM to generate train/test data. This enables the user to easily generate a dataset of their 2D material of interest. For this work, the model was trained on moiré twist angles of 0, 1, and 2 degrees of hBN encapsulated CrI3 bilayer and demonstrated a remarkable ability to generalize over unknown angles (1.5°, 2.5°, 3.5°, and 4°), thereby showcasing its robustness and versatility. The model precisely segments the atomic columns with a quantification accuracy of 99.27 ± 0.75%. This work highlights the potential of deep learning algorithms in automating and enhancing the analysis of 2D materials, paving the way for more efficient research and product development in the field.