The knowledge of atomistic structures is paramount for understanding a variety of chemical and physical processes. Systems with reduced symmetry, such as nanostructures and defects, are particularly challenging for atomistic structure prediction due to a large number of possible configurations. Materials interfaces can be characterized by electron microscopy in two dimensions; however, the three-dimensional atomic structure at the interface can be very difficult to determine, due to defects and disorder. In the present work, we aim to overcome this problem by combining atomistic modeling with electron microscopy image simulation and computer vision methods. Image segmentation, matching, and feature correspondences are among some of the most fundamental challenges in building robust vision systems.

In atomic resolution microscopy, these challenges are further complicated by both the high-degree of local, feature-level self-similarity in ordered regions, and the low SNR environments of disordered regions. To address issues pertaining to these challenges, we have developed a new feature space representation for atomic resolution microscopy images, that provides a pathway for segmenting, matching, and solving the correspondence problem for these images. Currently, the image features are extracted using a hand-crafted peak detection methods, which limits the types of microscopy images that this representation can be applied to. Recently, Facebook made its computer vision tools (DeepMask and SharpMask) open source. These tools provide deep learning approaches to automatic feature extraction which we believe could be leveraged to make our representation applicable to a wide variety of microscopy images in an automated setting.