Machine Learning-Driven Automated Detection and Classification of Defects in Atomic Resolution STEM Images

Conventional techniques, such as transmission electron microscopy (TEM), require manual identification of atomic columns, especially in regions with limited long-range order, making the process both labor-intensive and prone to error. Recent studies have integrated computer vision methods to automate defect detection in atomic-resolution STEM and TEM images. In response, we developed an unsupervised machine-learning model that autonomously detects defects in STEM images with minimal user input. This model demonstrates broad applicability across different atomic structures without requiring prior knowledge of bulk or defect configurations. Utilizing a Convolutional Variational Autoencoder (CVAE), our approach constructs an algorithm that robustly predicts the bulk structure, even in the presence of local perturbations. Additionally, it efficiently processes large STEM datasets, classifying images based on defect types with an accuracy of 98.7%. At an advanced stage, our CVAE-based algorithm is also capable of identifying subtle atomic-scale anomalies by discerning variations in pixel intensity, thereby enabling the differentiation of atoms according to their atomic number.