Integrating Chemical Insights and Experimental Parameters into Deep Learning Models Through Feature Engineering for Scanning Transmission Electron Microscopy Data

Automatic and high-throughput defect identification from scanning transmission electron microscopy (STEM) is crucial for establishing the structure-property relationships in materials. Identifying point defects in monolayer transition metal dichalcogenides (TMDs) is particularly challenging, as existing models often focus on a limited number of samples with distinct contrast features in their atomic columns.<br/><br/>In this study, we discovered that relying solely on imaging data, even advanced deep learning models struggle to accurately identify point defects. To address this limitation, we integrated chemical insights and experimental parameters into our Zernike feature encoder. This integration of domain-specific knowledge enables the classification of point defects across all 1H monolayer TMDs.<br/><br/>Our findings also reveal that with appropriate feature engineering, simple deep learning models like multi-layer perceptrons (MLPs) can achieve high accuracy in defect identification. This research has the potential to significantly expand current automated experimental workflows to a wider range of materials, enhancing the efficiency and scope of defect identification in various material systems.