Automatic identification of crystal structures and interfaces via artificial-intelligence-based electron microscopy

Andreas Leitherer, Byung Chul Yeo, Christian H. Liebscher & Luca M. Ghiringhelli 

npj Computational Materials  9: 179 (2023).

doi.org/10.1038/s41524-023-01133-1

Published online: 02 October, 2023

编辑概述

AI-STEM：自动识别晶体结构和界面

块材的不同晶体结构、表面和界面，以及纳米材料在先进材料性能的调控中发挥着关键作用。表征材料的缺陷或界面、局部晶格取向和畸变需要达到原子级的分辨率，而现代扫描透射电子显微镜通常能达到分辨率的需求，并能够以皮米的精度观察到原子的复杂排列。电子显微镜照片中的不同结构会展现出不同的明暗，但由于缺乏通用的自动分析工具，照片中可用的丰富信息还没有得到充分利用。近年来，大数据分析和人工智能（AI）在分析大型电子显微镜数据方面展示出巨大的潜力，可以揭示各种被忽视的影像特征，并且在各种数据集上得到了一些应用。经过适当训练的神经网络，如卷积神经网络（CNNs），被证明比其他机器学习方法更能准确地解决图像得分类问题，特别是在高通量任务中更为有效。在本工作中，来自德国马普学会弗里茨·哈伯研究所的Andreas Leitherer等人，提出了一种自动的、基于人工智能（AI）的方法，可用于在多晶材料的原子分辨率扫描透射电子显微镜（STEM）图像中准确识别关键特征。该方法是一个无监督的贝叶斯卷积神经网络，其中贝叶斯方法允许识别晶格对称性和晶体取向，而卷积方法则用于将图像分割成体区和界面区。研究人员使用理想晶格结构的模拟STEM图像进行训练，但训练的模型仍然可以推广到实验图像中。即使对于高噪声的单帧图像，该模型在块材区域的预测中也显示出很低的不确定性值。该方法不需要人工干预的情况下自动分析和分类STEM数据集，在研究晶体学属性方面具有巨大的潜力，为在原子水平上研究复杂的纳米结构铺平了道路。

Editorial Summary

AI-STEM: Automatic identification of crystal structures and interfaces

Different crystal structures, surfaces and interface play a key role in tailoring properties of advanced materials. Characterizing defects or interfaces, local lattice orientations and distortions requires atomic-level resolution. Modern electron microscopes can usually meet the resolution requirements and can be used to observe complex arrangements of atoms with picometer accuracy. Different structures in electron micrographs exhibit different shades of light and dark, but the rich information available in electron micrographs has not been fully exploited due to the lack of universal automated analysis tools. In recent years, big data analytics and artificial intelligence (AI) have shown great potential in analyzing large electron microscopy data to reveal various overlooked image features, and have found some applications on various data sets. Properly trained neural networks (NNs), such as convolutional neural networks (CNNs), have been shown to solve image classification problems more accurately than other machine learning methods, especially more efficiently than humans in high-throughput tasks. In this work, Andreas Leitherer et al. from the Fritz Haber Institute, Max Planck Society, Germany, proposed an automated, artificial intelligence (AI)-based method for accurately identifying key characteristics from atomic-resolution scanning transmission electron microscopy (STEM) images of polycrystalline materials. The method is an unsupervised Bayesian convolutional neural network, where the Bayesian method allows the identification of lattice symmetry and crystal orientation, while the convolutional method is used to segment the image into bulk and interface regions. The authors used the simulated STEM images of ideal lattice structures for training, but the trained model could be generalized to experimental images. Even for single-frame images with high noise, the model shows low uncertainty values in predictions of bulk crystal regions. This method has great potential to automatically analyze and classify the crystallographic properties of STEM data sets without the need for human intervention, paving the way for the autonomous study of complex nanostructures at the atomic level.