Metrology and characterization of defects in transition metal dichalcogenides using scanning tunneling microscopy enhanced with machine learning

Scanning tunneling microscopy (STM) has been shown to non-destructively determine the nature of many defects in TMDs. However, manual identification and counting of point defects in STM images is subjective and painstakingly slow, causing reported point defect densities to have large experimental uncertainty, preventing informed refinements of synthesis and controlled reduction in TMD point defect density. We have enabled atomic scale defect metrology of WSe₂ by enhancing STM with machine learning. Specifically, we leveraged recent machine learning advances in computer vision to automatically localize and identify defects. Eight types of defects were qualitatively identified and characterized by their apparent spatial extent. From this, we produced a dataset of 2979 instances of WSe₂ defects annotated with bounding boxes and associated labels. Two models were trained on our dataset and predictions were compared. Transfer learning, where the final layer of the classifier was replaced and re-trained to improve predictive accuracy, was used to adapt a pre-trained ResNet50 model to our dataset, while a Unet model was trained from scratch. Point defect densities in WSe₂ were determined with statistical significance using these trained detectors. Since defects are known to contribute towards electronic, structural, and optical properties in TMDs, our ML-enhanced STM can impact many 2D materials researchers.