Abstract:Scanning transmission electron microscopy (STEM) is an indispensable tool for atomic-resolution structural analysis for a wide range of materials. The conventional analysis of STEM images is an extensive hands-on process, which limits efficient handling of high-throughput data. Here we apply a fully convolutional network (FCN) for identification of important structural features of two-dimensional crystals. ResUNet, a type of FCN, is utilized in identifying sulfur vacancies and polymorph types of ${MoS_2}$ from atomic resolution STEM images. Efficient models are achieved based on training with simulated images in the presence of different levels of noise, aberrations, and carbon contamination. The accuracy of the FCN models toward extensive experimental STEM images is comparable to that of careful hands-on analysis. Our work provides a guideline on best practices to train a deep learning model for STEM image analysis and demonstrates FCN's application for efficient processing of a large volume of STEM data.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the high - throughput data processing challenges faced in scanning transmission electron microscope (STEM) image analysis, especially for the recognition of structural features in two - dimensional crystalline materials (such as molybdenum disulfide, MoS₂). Specifically, the paper mainly solves the following two key problems: 1. **Efficient identification of point defects and polymorphism**: - **Point defect identification**: The paper uses a deep - learning model (ResUNet) to identify sulfur vacancies in MoS₂. These point defects have an important impact on the electrical, chemical, and mechanical properties of the material. - **Polymorphism identification**: The paper also develops a model for identifying different stacking polymorphisms of MoS₂. These polymorphisms (such as 2H, 3R, and 1T stacking configurations) have a significant impact on the electronic band structure of the material. 2. **Improving the automation and accuracy of STEM image analysis**: - Traditional STEM image analysis relies on manual operation, which is time - consuming and inefficient and difficult to handle a large amount of high - throughput data. The paper trains a fully convolutional network (FCN), especially the ResUNet architecture, using simulated STEM images for training and introducing factors such as noise, aberration, and carbon contamination to improve the robustness and accuracy of the model in actual experimental images. - The paper demonstrates the effectiveness of the developed deep - learning model in processing experimental STEM images containing carbon contamination and other noise sources, proving that its performance is comparable to or even better than that of traditional manual analysis. ### Formula representation There are few formulas involved in the paper, but to ensure that all formulas can be presented accurately, the following are examples of relevant formulas in Markdown format: - **Loss function**: - Dice Loss: \[ \text{Dice Loss} = 1-\frac{2\times\text{TP}}{2\times\text{TP}+\text{FP}+\text{FN}} \] - Categorical Cross Entropy Loss: \[ \text{Categorical Cross Entropy Loss}=-\sum_{i = 1}^{N}y_i\log(\hat{y}_i) \] ### Summary By applying deep - learning techniques, especially the ResUNet architecture, the paper has successfully achieved efficient identification of sulfur vacancies and polymorphism in MoS₂. This method not only improves the degree of automation of STEM image analysis but also provides new ideas and technical means for processing other two - dimensional crystalline materials.

STEM image analysis based on deep learning: identification of vacancy defects and polymorphs of ${MoS_2}$

Accurate atomic scanning transmission electron microscopy analysis enabled by deep learning

Single-image based deep learning for precise atomic defects identification

Single-Image-Based Deep Learning for Precise Atomic Defect Identification

Automatic Identification of Crystal Structures and Interfaces via Artificial-Intelligence-based Electron Microscopy

Deep Learning Analysis of Defect and Phase Evolution During Electron Beam Induced Transformations in WS2

Probing electron beam induced transformations on a single defect level via automated scanning transmission electron microscopy

Deep Learning of Atomically Resolved Scanning Transmission Electron Microscopy Images: Chemical Identification and Tracking Local Transformations

Deep-Learning Recognition of Scanning Transmission Electron Microscopy: Quantifying and Mitigating the Influence of Gaussian Noises

Robust multi-scale multi-feature deep learning for atomic and defect identification in Scanning Tunneling Microscopy on H-Si(100) 2x1 surface

Machine Learning Identification of Impurities in the STM Images

AtomGAN: unsupervised deep learning for fast and accurate defect detection of 2D materials at the atomic scale

A deep learning framework for automated molecular recognitions in scanning probe microscopy images.

Learning-based Defect Recognition for Quasi-Periodic Microscope Images

Exploring electron-beam induced modifications of materials with machine-learning assisted high temporal resolution electron microscopy

Unsupervised machine learning combined with 4D scanning transmission electron microscopy for bimodal nanostructural analysis

Detecting Atomic Scale Surface Defects in STM of TMDs with Ensemble Deep Learning

Intelligent Identification of MoS2 Nanostructures with Hyperspectral Imaging by 3D-CNN

Dynamic STEM-EELS for single-atom and defect measurement during electron beam transformations

Deep learning-based noise filtering toward millisecond order imaging by using scanning transmission electron microscopy

Disentangling multiple scattering with deep learning: application to strain mapping from electron diffraction patterns