Automated identification of unit cell defects in porous MOFs and zeolites

Abstract

Metal-Organic Frameworks (MOFs) and zeolites have been a crucial part of materials being used in catalysis, gas storage and separation due to their well-defined crystalline structures. Traditional methods of manual defect characterization using Transmission Electron Microscopy (TEM) images are time-consuming, subjective and require expert interpretation. This project aims to develop an automated identification approach to identify and classify unit cell defects in MOFs and zeolite structures with TEM images. This approach may facilitate the discovery of new zeolite structures. This study adopts an unsupervised clustering approach using DBSCAN which does not require labeled data and can identify known and potentially novel patterns. The methodology consists of RDF-guided patch extraction based on void detection, Gabor filtering to extract local texture features, dimensionality reduction using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding(t-SNE) for visualization. DBSCAN clustering is used to group the structurally similar patches. A cluster-averaged patch is manually selected and used for refined template matching where it will undergo a second round of feature extraction and clustering to obtain the result. The results show that filtered TEM images performed significantly better in terms of detection and classification of unit cells as compared to raw TEM images which are typically noisier and have lesser contrast to the image. It can effectively distinguish between MFI and MEL structure with the averaged cluster patches aligning closely with the reference result. However, challenges remain in handling very noisy images where feature extraction and clustering accuracy degrade. Future work will focus on optimizing additional feature engineering techniques to enhance classification accuracy and framework identification.