Quantitative Electron Microscopy of Zeolites Using Aberration Corrected (S)TEM and Machine Learning

Due to their porous frameworks, zeolites can function as molecular sieves and have a number of industrial applications. The presence of intergrowth defects between different zeolite topologies can strongly affect their material properties [1]. For instance, both ZSM-5 and the ZSM-5/ZSM-11 intergrowth can be used to catalyze the production of ethylbenzene, via the alkylation of benzene. It has been shown that the ZSM-5/ZSM-11 intergrowth is the superior catalyst of these two, with higher alkylation and catalytic stability compared to pure ZSM-5, which deactivates much faster [2]. This is attributed to the active sites of the ZSM-5/ZSM-11 intergrowth being more resistant to poisoning by impurities in the reaction mixture. It has also been shown that in the single catalytic reduction of NOx, the AlPO-18/chabazite intergrowth exhibited greater NOx and N2O selectivity as a catalyst than pure chabazite when aged, which was attributed to superior hydrothermal stability [3]. ZSM-5 and ZSM-11 correspond to structure codes MFI and to MEL, whereas AlPO-18 and chabazite correspond to structure codes AEI and CHA. Moving forward, this paper will use these codes when referring to the relevant zeolites. These examples illustrate how the presence of intergrowth defects affect the material properties of zeolites, which in turn impacts industrial catalytic applications. The ability to quantify how zeolite material properties are affected by intergrowth defects, by measuring relevant statistical data in intergrowth structures, may lead to new insights that influence the development of improved and higher-performance catalysts. Examples of such statistical data may include zeolite pore sizing distributions, the location and extent of different types of zeolite active sites and data that leads to the complete characterization of three-dimensional intergrowth structures. This paper reports the automated identification of intergrowth defects in (S)TEM images of zeolite structures. Historically, the identification of zeolite defects (such as intergrowths) from such images has been completed manually. Such a workflow is inefficient; it is time-consuming, potentially error-prone and laborious. As a solution, we propose the training of a semantic segmentation neural network called ZeoSegNN, which will identify intergrowth defects from (S)TEM images of zeolites. Augmented learning is utilized, wherein training and testing is performed using both computationally generated and experimental (S)TEM images of zeolite structures. Using computer-simulated zeolite images ensures a large volume of training data can be quickly generated, with experimental data providing important realistic augmentation data. Two versions of ZeoSegNN have been trained, to identify two different zeolite intergrowths. ZeoSegNN-v1 has been trained to identify the MFI/MEL intergrowth, whilst ZeoSegNN-v2 has been trained to identify the AEI/CHA intergrowth. Figure 1 illustrates how augmented learning has been used to train ZeoSegNN. To produce simulated (S)TEM images, a zeolite generator – ZeoPy – has been coded in Python. ZeoPy builds on the Atomic Simulations Environment (ASE) and MAZE libraries [4, 5] to generate large volumes of intergrowth structures and ‘perfect’ intergrowth-free structures at high-throughput. At present, ZeoPy can create AEI, CHA, MFI and MEL zeolites as well as the AEI/CHA and MFI/MEL intergrowths. Once these structures are produced, MULTEM [6] (software that uses the multislice algorithm [7, 8] to simulate (S)TEM imaging) was used to produce simulated images. Segmentation masks were also created for all generated (S)TEM images, with each mask consisting of labels corresponding to the atomic positions of any intergrowth defects present. Computer generated images comprised the training and testing datasets of ZeoSegNN. This is in part due to experimental imaging of zeolites being challenging, as these structures readily undergo radiolysis under an electron beam [9]. Through the use of an aberration-corrected JEOL ARM300CF microscope with a DE-16 direct electron detector, it is anticipated that low-dose atomic-resolution experimental (S)TEM images of pure zeolites and their intergrowths can be captured as described elsewhere for similarly beam sensitive samples [10] and incorporated into the training and testing datasets of the next iterations of ZeoSegNN. ZeoSegNN uses k-fold cross validation with an underlying U-net architecture [11] for training. For the most recent iteration of ZeoSegNN-v1, 20-fold cross validation with 25 epochs were used with 6530 computationally generated images to train the network, and 4587 unseen computationally generated images were used as tests. The mean and median Jaccard scored recorded were 0.90 and 0.96. For ZeoSegNN-v2, 5-fold cross validation with 100 epochs was used with 6156 computationally generated images to train the network, and 2060 unseen computationally generated images were used as tests. The mean and median Jaccard scores recorded were 0.73 and 0.82. We have developed a semantic segmentation neural network, ZeoSegNN, that can be used to identify intergrowth defects from (S)TEM images of AEI/CHA and MFI/MEL structures, using augmented learning. Schematic diagram illustrating the use of augmented learning to train/test ZeoSegNN. (a) ZSM-5/ZSM-11 pentasil intergrowth model generated in ZeoPy, (b) MULTEM simulation of model, (c) Segmentation mask of the intergrowth defect. (a) Unseen AEI/CHA intergrowth testing image for ZeoSegNN-v2, (b) ground truth intergrowth position segmentation mask, (c) prediction from by ZeoSegNN-v2 for the intergrowth position. This example has a Jaccard score of 0.825.