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

Area of Science:

• Materials Science
• Condensed Matter Physics
• Nanotechnology

Background:

• Point defects in 2D materials like MoS₂ significantly influence electronic and optoelectronic properties.
• Accurate defect identification is crucial for understanding defect physics and optimizing device performance.

Purpose of the Study:

• To develop and validate a deep learning pipeline for automated defect segmentation and classification in MoS₂ using STM images.
• To integrate image analysis with physics-based modeling for comprehensive defect characterization.

Main Methods:

• Acquisition of high-resolution STM images of monolayer MoS₂.
• Application of the Segment Anything Model (SAM) for automatic defect region segmentation.
• Classification of segmented regions using a convolutional neural network (CNN) trained on augmented data.
• Validation against manual annotations and support from density functional theory (DFT) calculations.

Main Results:

• The deep learning pipeline achieved 95.06% classification accuracy on a dataset of 198 samples.
• The model demonstrated robustness and effectiveness despite limited data.
• DFT calculations confirmed the presence of localized mid-gap states associated with sulfur vacancies, consistent with STM observations.

Conclusions:

• The combined approach of SAM segmentation, CNN classification, and DFT modeling offers a powerful method for quantifying defect populations in MoS₂.
• This data-driven and physics-based strategy can accelerate defect characterization in 2D materials.