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

Deep learning techniques based on fully convolutional networks (FCNs) have revolutionized image recognition in fields ranging from medical diagnosis to facial recognition. The ability of FCNs to identify objects in images opens new opportunities for accessing the underlying information in atomic-resolution images obtained using scanning transmission electron microscopy (STEM). The properties of two-dimensional transition metal dichalcogenides (2D TMDCs) are strongly influenced by atomic defects such as vacancies and substitutional dopants, but high-precision characterization of single-atom defects remains challenging because 2D materials are irradiation sensitive, produce low scattering signals, and require low-voltage imaging modes. While identifying defects by hand is possible, it severely limits our ability to process large quantities of atoms and obtain large-scale statistics. By employing deep learning techniques, we quickly identify and classify various defect species, including metal substitutions, chalcogen vacancies, and chalcogen substitutions. This approach lets us observe changes in atomic separations induced by these defects with sub-picometer precision.