A study published in Molecules and led by researchers from the Changchun Institute of Optics, Fine Mechanics and Physics (CIOMP) of the Chinese Academy of Sciences demonstrated how deep learning can streamline the identification of atomic-scale defects in molybdenum disulfide (MoS₂), a promising two-dimensional (2D) material for next-generation electronics.

Researchers addressed a critical challenge in materials science: Manually locating and classifying point defects in scanning tunneling microscopy (STM) images is time-consuming and prone to human error. They developed a hybrid approach combining the Segment Anything Model (SAM) for defect localization with a convolutional neural network (CNN) for classification. 

The deep learning pipeline was trained on just 198 augmented STM images of MoS₂ grown on gold substrates, achieving 95.06% accuracy in distinguishing sulfur vacancies (V_S) from interstitials (A_S) and Moiré patterns. 

The key to the approach’s success was preprocessing STM images to suppress noise while preserving defect signatures. SAM first segmented candidate regions which were then classified by the CNN. Density functional theory simulations showed that sulfur vacancies created localized mid-gap states visible in STM as hexagonal or triangular features. 

The performance of the SAM-CNN integrated approach surpassed traditional tools like OpenCV, particularly in handling small datasets, a limitation in experimental materials science.

This work offers a scalable solution for defect analysis in 2D systems. The proposed approach could accelerate quality control in semiconductor fabrication or the study of defect-engineered quantum materials. Future improvements may involve expanding the training dataset to cover more defect types and material systems.