Development of Automated Methods for Extraction of Structural Information from Microscopy Data
Ovchinnikov, Oleg Sergeevich
Over the last few decades the capabilities of atomic-resolution imaging techniques such as Scanning-Transmission-Electron-Microscopy (STEM) have greatly increased. The data acquisition rate of modern of atomic-resolution imaging techniques has far outpaced the ability to analyze the data. Many times the operator chooses only a few images to analyze while discarding the rest. Most current algorithms for image analysis require user input on multiple steps or significant training on the material creating a bottleneck. Automated methods would allow for the analysis of all the collected data, allowing for statistical studies about the material properties and defect concentration and distribution. Automated methods could also provide real-time information to the operator, helping guide the operator to areas of interest. In this thesis, we develop a set of methods for the automated analysis of atomic-resolution images and test them on STEM data. By using only the information collected alongside the image, a rapid algorithm based on a whole-image transform can identify atomic columns in 1 sec or less on an average desktop-style computer. Subsequently using the atomic-columns, a rapid, graph-theory-based approach is developed that can reliably identify (also in 1 sec or less) a variety of structural defects using the found atomic column positions. This approach is able to detect defects with no prior knowledge of the lattice type(s) present and no requirement to continuity thereof. Results indicate that substitutional defects, which often result in local lattice disturbance, can also be successfully identified with this approach. Extracted information can provide input for theoretical calculations such as Density Functional Theory, which can, for example, determine not only the distribution of atomic columns within the image plane, but also find full 3D coordinates of atoms via total-energy minimization. As a proof-of-principle study, this type of analysis/DFT hybrid approach was pursued using a STEM image of angle-missmatched bilayer, 3D structure was revealed and nanoscale ripples present in the system were quantified. Furthermore this relaxed structure was used as proof-of-principle for an iterative image-analysis-DFT-simulation loop that allowed for identification of defects in the crystal lattice, enabling us to perform data-to-simulation comparison on a quantitatively new level.