無監(jiān)督機(jī)器學(xué)習(xí)：原子分辨圖像中的缺陷檢測

晶體缺陷在決定材料特性方面起著關(guān)鍵作用。例如，二維材料中的點(diǎn)缺陷（如空位和間隙）可以為晶格引入應(yīng)變并改變電子特性；線缺陷（如位錯(cuò)）和平面缺陷（如?3D?材料中的晶界和孿晶界）可以決定它們的機(jī)械性能。晶體缺陷的研究需要原子尺度分辨率的晶體缺陷結(jié)構(gòu)，例如來自掃描透射電子顯微鏡?(STEM)?圖像的晶體缺陷。目前將機(jī)器學(xué)習(xí)?(ML)?應(yīng)用于電子顯微鏡的趨勢需要“人工智能晶體學(xué)家”，它可以自動(dòng)檢測、識別和分析結(jié)構(gòu)特性。基于監(jiān)督學(xué)習(xí)的方法可以對缺陷進(jìn)行檢測和分類，它使用來自實(shí)驗(yàn)或模擬的人工標(biāo)記數(shù)據(jù)來訓(xùn)練模型。這些經(jīng)過訓(xùn)練的模型通常僅限于訓(xùn)練數(shù)據(jù)集中的晶體結(jié)構(gòu)，不能輕易推廣到其他晶體結(jié)構(gòu)。目前，缺乏一種可靠的無監(jiān)督機(jī)器學(xué)習(xí)方法來檢測?STEM?圖像中的晶體缺陷。
Fig. 1 Results of defect detection in bilayer Mo0.91W0.09Te2.?
來自美國橡樹嶺國家實(shí)驗(yàn)室的Lupini和Yueming?Guo研究團(tuán)隊(duì)，報(bào)導(dǎo)了一類支持向量機(jī)（OCSVM）來進(jìn)行晶體缺陷檢測，這種方法在工業(yè)中廣泛用于異常值檢測，例如制造和網(wǎng)絡(luò)安全分析中的快速故障診斷。他們認(rèn)為晶體缺陷的檢測可以轉(zhuǎn)化為一類分類背景下的異常值檢測問題。作者引入了兩種分割方案來滿足具有一類主要數(shù)據(jù)點(diǎn)的要求：第一種方案適用于廣泛的晶體結(jié)構(gòu)，并對晶體對稱性施加了限制；第二種方案裁剪傅立葉濾波圖像，使每個(gè)裁剪圖像都以原子列的位置為中心，消除了晶體對稱性的限制。為了確保每個(gè)子圖像中的特征與平移無關(guān)，他們將帕特森函數(shù)作為特征提取描述符引入了兩種方案，使其僅依賴于原子間向量（不包括它們的方向符號）。作者報(bào)導(dǎo)的這類方法可用于快速篩選大量?STEM?圖像以檢測晶體缺陷，還可以改善來自光束敏感材料（例如金屬有機(jī)框架化合物?(MOF)）的噪聲原子圖像的晶體結(jié)構(gòu)重建。該文近期發(fā)表于npj Computational Materials?7:?180(2021)。
Fig. 2 Illustration of how Patterson mapping can greatly reduce the effect of arbitrary shifts in the fifirst demonstration.??
Editorial Summary
Unsupervised?ML:?Defect detection in atomic-resolution images
Crystallographic defects play a key role in defining materials properties. For example, point defects in 2D materials such as vacancies and interstitials can introduce strain to the lattice and modify the electronic properties; line defects like dislocations and planar defects such as grain boundaries and twin boundaries in 3D materials can define their mechanical properties. Studies of crystallographic defects require structural knowledge of crystallographic defects at atomic-scale resolution,?such as crystallographic defects from scanning transmission electron microscopy (STEM) images.?The current trend of applying machine learning (ML) to electron microscopy demands an ‘AI crystallographer’ that can automatically detect, identify, and analyze structural properties. Previous work demonstrated the detection and classification of defects based on supervised learning methods where human-labeled data from experiments or simulations are used to train the models.?These trained models are usually limited to the crystal structures in the training datasets and cannot easily be generalized to other crystal structures. Currently, a reliable method of unsupervised ML for detection of crystallographic defects in STEM images is lacking.
Fig. 3 Details of OCSVM analysis.
A group led by Lupini and Guo?from?Oak Ridge National Laboratory, USA, proposed the application of the one-class support vector machine (OCSVM), which has been widely used for outlier detection in industry (such as fast fault diagnosis in manufacturing and cyber security analyses), to the detection of crystallographic defects. Such detection of crystallographic defects can be turned into a problem of outlier detection in the context of one-class classification. The authors introduced two schemes of segmentation to satisfy the requirement of having one major class of data points. The first scheme is applicable to a wide range of crystal?structures although it imposes limitations to the crystal symmetry. To remove the limitations, they?developed a second scheme where they?crop the Fourier filtered images and each cropped image is centered at the position of an atom column. To make sure that the features in each subimage are independent of translation, they introduced the application of the Patterson function?as a feature extraction descriptor to both schemes, which only depends on the interatomic vectors (excluding the signs of their directions). In?this article, the as-reported approach by authors can be applied to rapidly screen a large number of STEM images to detect crystallographic defects, which may be beneficial when analyzing?in situ?data. Another application is to improve the crystal structure reconstruction from noisy atomic images of beam-sensitive materials such as metal-organic frameworks (MOFs). This?article was recently?published in npj Computational Materials 7: 180?(2021).
Fig. 4 Demonstration of detecting different domains separated by the twin boundary in a ZrO2?nanoparticle.
原文Abstract及其翻譯
Defect detection in atomic-resolution images via unsupervised learning with translational invariance?(利用具有平移不變性的無監(jiān)督學(xué)習(xí)在原子分辨率圖像中檢測缺陷)
Yueming Guo,?Sergei V. Kalinin,?Hui Cai,?Kai Xiao,?Sergiy Krylyuk,?Albert V. Davydov,?Qianying Guo?&?Andrew R. Lupini?
Abstract?Crystallographic defects can now be routinely imaged at atomic resolution with aberration-corrected scanning transmission electron microscopy (STEM) at high speed, with the potential for vast volumes of data to be acquired in relatively short times or through autonomous experiments that can continue over very long periods. Automatic detection and classification of defects in the STEM images are needed in order to handle the data in an efficient way. However, like many other tasks related to object detection and identification in artificial intelligence, it is challenging to detect and identify defects from STEM images. Furthermore, it is difficult to deal with crystal structures that have many atoms and low symmetries. Previous methods used for defect detection and classification were based on supervised learning, which requires human-labeled data. In this work, we develop an approach for defect detection with unsupervised machine learning based on a one-class support vector machine (OCSVM). We introduce two schemes of image segmentation and data preprocessing, both of which involve taking the Patterson function of each segment as inputs. We demonstrate that this method can be applied to various defects, such as point and line defects in 2D materials and twin boundaries in 3D nanocrystals.
Fig. 5 Demonstration of applying Patterson function to provide translation-invariant features in segmented subimages.
摘要?現(xiàn)在可以通過像差校正的掃描透射電子顯微鏡?(STEM)?以原子分辨率對晶體缺陷進(jìn)行高速成像，并有可能在相對較短的時(shí)間內(nèi)或通過可以持續(xù)很長時(shí)間的自主實(shí)驗(yàn)獲得大量數(shù)據(jù)。為了能夠以有效的方式處理數(shù)據(jù)，需要對?STEM?圖像中的缺陷進(jìn)行自動(dòng)檢測和分類。然而，與人工智能進(jìn)行物體檢測和識別任務(wù)一樣，從?STEM?圖像中檢測和識別缺陷具有挑戰(zhàn)性。此外，處理具有許多原子和低對稱性的晶體結(jié)構(gòu)是有難度的。以前用于缺陷檢測和分類的方法是基于監(jiān)督學(xué)習(xí)，需要人工標(biāo)記數(shù)據(jù)。在本工作中，我們開發(fā)了一種基于一類支持向量機(jī)?(OCSVM)?的無監(jiān)督機(jī)器學(xué)習(xí)缺陷檢測方法。我們介紹了圖像分割和數(shù)據(jù)預(yù)處理兩種方案。這兩種方案都涉及將每段的帕特森函數(shù)作為輸入。我們證明了這種方法可以應(yīng)用于各種缺陷，例如?2D?材料中的點(diǎn)和線缺陷以及?3D?納米晶體中的孿晶界。
轉(zhuǎn)自：npj計(jì)算材料學(xué)