Crystal Structure Prediction from Atomic-Resolution Electron Microscopy Images using Deep Learning

Accurate identification and classification of crystal structures, in many technological applications, is one of the first steps towards extracting useful information from atomic-resolution images. However, identifying the structural information from these images using traditional real-space approaches, such as finding local intensity maxima and template matching is challenging due to the lack of robustness of the image analysis methods. To this end, the development of a fast and automated structure prediction tool using AI/ML will play an important role in improving the prediction accuracy. Here, we discuss a proof-of-concept study to demonstrate performance of a deep neural network leveraging the transfer learning approach to learn different crystal phases from real space high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) images. To this end, we take around 200,000 simulated real space HAADF-STEM images, collected and preprocessed from Atomagined dataset (https://github.com/MaterialEyes/atomagined), for the deep learning training. Finally, we test the performance and accuracy of the network on both simulated test dataset and experimental images curated from existing literature.