A Smart Automated TEM Facility for Large Scale Analysis of Atomic Structure and Chemistry

从航空航天和汽车工业，到半导体，再到外科植入物，先进材料是众多关键现代技术的核心。透射电子显微镜（TEM）是先进材料研究的关键技术，因为它提供了两个重要的原子信息：首先，原子的位置可以通过研究样品中电子的弹性散射来确定，其次，材料结构中原子位置的化学组成可以通过对能量向样品的非弹性转移的光谱研究（通过直接能量损失或通过特征x射线的检测）来恢复。这两条信息支撑了一个巨大的研究领域，探索材料的微观结构和宏观性能之间的关系。随着纳米技术和量子器件的发展，发现对单个原子最精确理解的能力是支持先进材料研究的设施的基本能力。该项目的目的是首次开发一种分析TEM，不仅提供尖端的光谱性能，而且还以人工智能和自动化工作流程为核心设计。第一个目标将通过结合最新一代的x射线探测器和光谱仪来实现，以提供化学灵敏度和精度的数量级改进。这种能力对于研究小到单个原子缺陷的器件以及在原子分辨率下对大面积进行有效分析是必不可少的。为了实现人工智能（AI）辅助实验，该项目将解决一些技术挑战：1。将开发TEM的计算机控制，允许计算机自动调整样品级和光束以处理感兴趣的特定区域，并执行自动调整实验参数，以实现详细的高分辨率成像和基于纳米区域的衍射分析，而不需要连续的操作员交互。识别感兴趣区域的机制将利用各种机器学习（ML）方法来分割低分辨率数据，这些数据可能来自TEM中的快速大面积扫描，也可能是光学成像、扫描探针显微镜、扫描电子显微镜或光学方法进行非原位分析的结果。人工智能训练将允许显微镜控制计算机以与人类操作员相同的方式在实验结果之间建立功能关系，从而更快，更系统地识别新特征。我们提出的新的智能自动化TEM （AutomaTEM）提供了通过自动化工作流程分析轻松访问的数据量至少增加一个数量级的机会。感兴趣的特征将通过用户定义的参数或通过人工智能识别集体数据中的重要特征来确定。这将允许有意义的统计数据收集的大小，形状，原子结构，组成，电子行为的潜在数百或数千个区域在一个给定的样品。这反过来将使纳米结构的非均质性和材料的结构-性能关系有更全面的了解。

Advanced materials lie at the heart of a huge number of key modern technologies, from aerospace and automotive industries, to semiconductors through to surgical implants. The transmission electron microscope (TEM) is a key enabling technology for advanced material research because it offers two important pieces of atomic information: firstly the location of atoms can be determined from studies of elastic scattering of electrons by the sample, and secondly the chemical composition of atomic sites within the materials structure can be recovered from spectroscopic studies of the inelastic transfer of energy to the sample (either from direct energy loss or by the detection of characteristic X-rays). These two pieces of information underpin a huge research area exploring the relationship between materials microscopic structure and the macroscopic properties it exhibits. With the drive towards nanotechnologies and quantum devices the ability to discover the most precise understanding of individual atoms is an essential capability for facilities supporting research of advanced materials.The aim of the project is to develop, for the first time, an analytical TEM that not only offers cutting edge spectroscopy performance but which also is designed with artificial intelligence and automated workflows at its core. The first goal will be achieved through the incorporation of the latest generation of X-ray detectors and spectrometers to provide order of magnitude improvements in chemical sensitivity and precision. This capability is essential for the move to studying devices as small as a single atomic defect as well as for efficient analysis of large areas at atomic resolution. To achieve artificial intelligence (AI)-assisted experiments the project will tackle a number of technical challenges:i. Computer control of the TEM will be developed, allowing the computer to automatically adjust the sample stage and beam to address specific regions of interest and perform auto-tuning the experimental parameters to achieve detailed high resolution imaging and diffraction based analysis of nanometric regions without the need for continuous operator interaction.ii. The mechanism to identify regions of interest will utilise the full range of machine learning (ML) approaches to segment lower resolution data, which might come from fast large-area scanning in the TEM or be the result of ex-situ analysis by optical imaging, scanning probe microscopies, scanning electron microscopy or optical approaches to name but a few. iii AI training will allow the microscope control computer to build functional relationships between experimental results in the same way a human operator does, allowing faster and more systematic identification of novel features.Our proposed new smart automated TEM (AutomaTEM) offers the opportunity to gain at least an order of magnitude increase in the volume of data that is readily accessible through automated workflow analysis. Features of interest will be determined either through user-defined parameters or through the AI identification of significant features in the collective data. This will allow meaningful statistics to be gathered about the size, shape, atomic structure, composition, electronic behaviour of potentially hundreds or thousands of regions in a given sample. This in turn will enable more complete understanding of nanostructure heterogeneity and structure-property relationships in materials.