| Three-dimensional characterization of defects is an essential step in the engineering of point defects to control the type, concentration, and spatial distribution of defects with nanometer-scale resolution to design materials with new functions and properties. High-resolution electron microscopy is becoming a general-purpose tool for characterizing several point defects with single-defect sensitivity and sub-unit cell spatial resolution in all three dimensions. Detectable defects include substitutional impurities, interstitial impurities, self-interstitials, and impurity-containing defect complexes. However, all the defects so far imaged at the single-defect level often increase the local electron scattering by 50% or more, and 3D imaging of defects that change column intensity less than 10%, such as single vacancies was only reported very recently.;In the first part of this thesis, we demonstrated an approach for three-dimensional imaging of single vacancies using high precision quantitative high-angle annular dark-field Z-contrast scanning transmission electron microscopy. Vacancies are identified by both the reduction in scattered intensity created by the missing atom and the distortion of the surrounding atom positions. Vacancy positions are determined laterally to a unique lattice site in the image and in depth to within one of two lattice sites by dynamical diffraction effects. 35 single La vacancies are identified in images of a LaMnO3 thin film sample. The vacancies are randomly distributed in depth and correspond to a La vacancy concentration of 0.79%, which is consistent with the level of control of cation stoichiometry within our synthesis process (∼1%) and with the equilibrium concentration of La vacancies under the film growth conditions. This method can be extended to detect other defects including impurities and defect clusters and these results represent a step forward in characterizing point defects in materials one at a time, at atomic resolution, matching our current capabilities in materials simulation and our growing control over defect distributions in synthesis.;In the second parts of this thesis, I adapted the cut-edge Poisson denoising and machine learning algorithm into the four-dimensional STEM, which could potentially detect point defects that are undetectable by traditional STEM. We demonstrate that the iterative BM4D Poisson denoising algorithm could recover most of the image features corrupted by Poisson noise and increase the PSNR most. We also demonstrate that the convolutional neural network (VGG-16), trained on simulated PACBED data set, could accurately predict TEM sample thickness with > 99% accuracy within 100 nm with 2 nm thickness step. |
