Machine learning assisted fast optical identification of 2D materials

作     者：He, Junqi Zhou, Yi Zhang, Pan Zhai, Huifei 

作者机构：China Jiliang Univ Coll Sci Hangzhou 310018 Zhejiang Peoples R China Arizona State Univ Sch Mol Sci Tempe AZ 85287 USA 

出 版 物：《APPLIED SURFACE SCIENCE》 (Appl Surf Sci)

年 卷 期：2025年第686卷

核心收录：

学科分类：07[理学] 0805[工学-材料科学与工程（可授工学、理学学位）] 0703[理学-化学] 0702[理学-物理学] 

基　　金：Natural Science Foundation of Shaanxi Province [2022JM-028] Key Research and De-velopment Program of Shaanxi Province [2022GY-356] 

主　　题：2D materials Machine learning YOLO Clustering algorithm Instance segmentation 

摘      要：Machine learning has driven research and development in the field of materials. We investigate the thickness identification of few-layer tungsten ditelluride (WTe2) flakes based on optical microscopy images, combined with machine learning, and found a fast method for identifying 1- to 4-layers WTe2. We only use a single channel to obtain features. Green channel values of color are used to obtain the optical contrast. Unsupervised learning algorithms, including nearest neighbor assignment clustering, k-means clustering, and Mean-Shift clustering, are employed to extract data features and then build a thickness database of few-layer WTe2. Our results show that via machine learning analysis, the correlation between layer number counts and optical contrast values can be derived, and the layer-number-dependent optical reflection can be well described by the Beer-Lambert law. This provides a fast and reliable way to determine the thickness of WTe2 flakes, which will greatly improve the efficiency of identifying atomically thin samples. After obtaining the classification and labeling, we use the deep learning YOLOv8 model to segment optical microscope images. We have enhanced the model s accuracy by improving the algorithm, resulting in excellent generalization ability and robustness. Our proposed method is faster, more accurate, and more convenient than usual methods. This method can also reduce costs and time, avoiding the hassle of batch measurement.