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

作　　者：Danpeng CHENG Wuxin SHA Zuo XU Shide LI Zhigao YIN Yuling LANG Shun TANG Yuan-Cheng CAO 

机构地区：[1]State Key Laboratory of Advanced Electromagnetic Engineering and Technology,School of Electrical and Electronic Engineering,Huazhong University of Science and Technology,Wuhan,430074,China [2]School of Computer Science and Technology,Huazhong University of Science and Technology,Wuhan,430074,China [3]CITIC Dicastal Co.,Ltd.,Qinhuangdao,066011,China

出　　处：《Science China(Information Sciences)》2023年第6期98-110,共13页中国科学（信息科学）（英文版）

基　　金：supported by National Key R&D Program of China (Grant No.2022YFB2404303);National Natural Science Foundation of China (Grant Nos.52107224,52077096);State Grid Corporation of China (Grant No.520626210064);China Postdoctoral Science Foundation (Grant No.2019M662612)。

摘　　要：The extraction of atomic-level material features from electron microscope images is crucial for studying structure-property relationships and discovering new materials.However,traditional electron microscope analyses rely on time-consuming and complex human operations;thus,they are only applicable to images with a small number of atoms.In addition,the analysis results vary due to observers’individual deviations.Although efforts to introduce automated methods have been performed previously,many of these methods lack sufficient labeled data or require various conditions in the detection process that can only be applied to the target material.Thus,in this study,we developed AtomGAN,which is a robust,unsupervised learning method,that segments defects in classical 2D material systems and the heterostructures of MoS_(2)/WS_(2)automatically.To solve the data scarcity problem,the proposed model is trained on unpaired simulated data that contain point and line defects for MoS_(2)/WS_(2).The proposed AtomGAN was evaluated on both simulated and real electron microscope images.The results demonstrate that the segmented point defects and line defects are presented perfectly in the resulting figures,with a measurement precision of 96.9%.In addition,the cycled structure of AtomGAN can quickly generate a large number of simulated electron microscope images.

关 键 词：deep learning generative adversarial network defect detection atomic resolution 2D materials 

分 类 号：TB34[一般工业技术—材料科学与工程] TP18[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]