Research

Lab of Artificial Intelligence in Biomedical Imaging

Atomic Scale Imaging

Scanning Transmission Electron Microscope (STEM) image analysis using structural information

Scanning Transmission Electron Microscope (STEM) has become one of the most powerful techniques for local structural analysis beyond the spatial resolution limits of X-ray based characterizations. Based on an understanding of these structures, it is important to link the relationship between structures and properties. However, STEM obtains data in the form of scanning a sample of an electron beam, so if the scan speed is fast or the signal from the atom is very weak, the signal-to-noise ratio is lowered. Noise in STEM images can be modeled mainly as Poisson and Gaussian distributions. These noises make it difficult to analyze STEM images and cause problems with poor accuracy. Therefore, in this study, noise in STEM images is removed using Generative Adversary Networks (GAN) among deep learning architecture. Furthermore, we developed a program that analyzes the characteristics of the atomic structure through precise peak coordinates using segmentation.

Automated classification of point-defects in STEM image using CNNs

Atomic defects such as vacancies, edges, and grain boundaries are omnipresent in two-dimensional (2D) materials. Recently, a variety of researches focus on unveiling the relation between atomic defect structures and electronic, magnetic, or thermoelectric properties of 2D materials. These atomic imperfections often introduce new functionality in 2D materials and dictate such unprecedented properties. Therefore, comprehending the roles of defects offers opportunities to tune and modulate for new functionalities and applications. In defect structures, zero-dimensional(same point) defect analysis, which is an atomic-level defect, can be observed in higher-dimensional defects. In addition, the development of STEM (scanning transmission electron microscopy) enabled picometer-level point defect analysis. The point dimensional analysis of 2D materials using conventional STEM images was performed by comparing the intensity profile with the simulated image. (However, this method is somewhat tedious and inaccurate because it is not objective.) Therefore, we are researching a method for analyzing defects in STEM images using convolutional neural networks.