Using Artificial Intelligence to Understand Zeolite Catalysts

电子显微镜为在原子尺度上研究材料的局部结构提供了一种独特的工具。然而，一个主要的挑战在于将大量测量的属性与高度选择性的结构数据相关联。一种方法是利用快速电子探测器来获取数百万张图像的大型数据集，并开发基于深度学习的自动分析工具来分析这些图像。本项目的总体目标是研究涉及多相催化的沸石中的缺陷结构。该项目将使用最新发展的快速直接电子探测器，用于透射电子显微镜的低剂量成像和神经网络训练的特定缺陷结构的模式识别。这些研究的潜在影响是更好地了解工业赞助商感兴趣的催化过程，并更好地了解催化性能和局部结构之间的关系。最初，该项目将使用以kHz帧速率运行的新探测器来记录包含许多缺陷结构TEM图像的大型数据集。与此同时，该项目将开发基于卷积神经网络的机器学习，以构建适合分析包含数百万图像的大型数据集的图像分析工具。卷积神经网络将使用已知缺陷结构的模拟数据针对各种电子剂量预算和其他成像条件进行训练。然后将其用于分析实验数据，以获得有关缺陷类型的有意义的统计数据。拟议的研究在很大程度上依赖于电子物理科学成像中心提供的独特仪器。具体而言，将使用一种新的高速直接电子探测器，在12位计数模式下以超过2KHz的帧率运行，以采集低剂量数据。在所有上述目的和目标中，将有必要确保所开发的方法对于低剂量数据采集是稳健的，因为沸石已知是辐射敏感的，并且确保电子束诱导效应最小化。最初将研究纯沸石，但该项目也将扩展到研究金属负载的沸石和这些系统中的缺陷和纯材料之间的比较将提供对金属负载的机制和结构后果的见解。最后，将在工业赞助商实验室测量催化数据，以尝试将催化性能与一系列加载和未加载样品中存在的缺陷的性质和密度相关联。该项目福尔斯属于EPSRC能源，人工智能和机器人技术以及物理科学研究领域。该项目由约翰逊Matthey plc通过iCase计划资助。

Electron microscopy provides a unique tool for studying the local structure of materials at the atomic scale. However, a major challenge lies in correlating bulk measurements of properties with highly selective structural data. One approach is to take advantage of fast electron detectors to acquire large data sets of millions of images and to develop automated analysis tools based on deep learning to analyse these. The overall aim of this project is to study defect structures in zeolites involved in heterogeneous catalysis. This project will use recent developments in fast direct electron detectors for transmission electron microscopy for low dose imaging and neural networks trained for pattern recognition of specific defect structures. The potential impact of these studies is a better understanding of catalytic processes of interest to the industrial sponsor and an improved understanding of the relationships between catalytic performance and local structure.Initially the project will use new detectors operating at kHz frame rates to record large datasets containing many TEM images of defect structures. In parallel the project will develop the use of machine learning based on convolution neural networks to build image analysis tools suitable for analysing large data sets containing millions of images. A convolutional neural network will be trained using simulated data of known defect structures for various electron dose budgets and other imaging conditions. This will then be used to analyse the experimental data to gain meaningful statistics on defect types. The research proposed relies heavily on unique instrumentation available at the electron Physical Sciences Imaging Centre. Specifically, a new high speed direct electron detector operating at a frame rate in excess of 2KHz in 12 bit counting mode will be used to acquire low dose data. Within all of the above aims and objectives it will be necessary to ensure that the methods developed are robust to low dose data acquisition as zeolites are known to be radiation sensitive and to ensure that electron beam induced effects are minimised. Initially pure zeolites will be studied but the project will also be extended to study metal loaded zeolites and comparisons between defects in these systems and the pure materials will provide insights into the mechanisms and structural consequences of metal loading. Finally catalytic data will be measured at the industrial sponsors laboratories to attempt to correlate catalytic performance with the nature and densities of defects present across a range of loaded and unloaded samplesThe project falls within the EPSRC energy, Artificial Intelligence and Robotics and physical sciences research areasThe project is funded by Johnson Matthey plc through the iCase initiative.