Point Defect Detection and Classification in MoS2 Scanning Tunneling Microscopy Images: A Deep Learning Approach

Abstract

Point defects in two-dimensional materials such as MoS2 can critically impact their electronic and optoelectronic properties. Precise identification of these defects is essential for understanding defect physics and device performance. In this work, we acquire high-resolution scanning tunneling microscopy (STM) images of monolayer MoS2 and apply the Segment Anything Model (SAM) to automatically segment possible defect regions in the STM images. Each segmented region is then classified by a convolutional neural network (CNN) architecture into defect categories. This deep learning pipeline is trained on augmented STM image data and evaluated against manual annotations. The model achieves a classification accuracy of 95.06% on a modest dataset comprising merely 198 samples, demonstrating its robustness despite limited data availability. We also perform density functional theory (DFT) calculations of representative defect structures to support interpretation of the STM features. Charge density isosurfaces of the DFT models reveal localized mid-gap states associated with sulfur vacancies, consistent with STM observations. The integration of SAM segmentation, CNN classification, and DFT modeling provides a comprehensive approach to quantify defect populations in MoS2. These results show the potential of combining data-driven image analysis with physics-based modeling to accelerate defect characterization in 2D materials.