Deep Learning Enabled Measurements of Single-Atom Defects in 2D Transition Metal Dichalcogenides with Sub-Picometer Precision

Abstract

Pushing the limits of electron microscopy requires both technical development in instrumentation and the utilization of new approaches for data extraction. In our work, we apply deep learning techniques based on convolutional neural networks (CNN), which have already revolutionized image recognition in fields such as medical diagnosis, weather forecasting, and facial recognition, to atomic-resolution images obtained using aberration-corrected scanning transmission electron microscopy (STEM) [1]. By utilizing the strong pattern recognition ability and adaptability of CNN, we provide new opportunities to access the underlying information in large volumes of atomic-resolution images of two-dimensional transition metal dichalcogenides (2D TMDCs), whose properties have been strongly influenced by atomic defects such as vacancies and substitutional dopants. Yet, high-precision characterization of single-atom defects remains challenging because 2D materials are irradiation sensitive, produce low signals, and require low-voltage imaging modes. Typical imaging modes struggle to provide sufficiently high signal-to-noise ratio (SNR) and resolution to measure the structure of single-atom defects without inducing electron beam damage. Here, we utilize CNN to identify atomic defects in 2D TMDCs such as Mo1-xWxTe2 and WSe2-xTex and use the resulting data to tackle the above-mentioned characterization challenges in 2D materials. Using these techniques, we are able to measure local lattice strains induced by substitutional dopants with subpicometer precision, which have been traditionally difficult to extract without the aid of CNN.