Significant improvements have been made in pipelines for the acquisition andstorage of data from scanning electron imaging experiments. Within these
data contain valuable information such as structure, defects and compositions of
samples that must first be inverted via an expensive numerical process. A
typical scanning CDI imaging setup records measurements in 4-dimensions -- 2
from scanning and 2 from scattering dimensions, and with countless microscopy
experiments and continuous infrastructure upgrades, the computational demand for
facilitating the inversion process and analyzing the results grow at an
exponential pace.
A numerical study of the advantages of incorporating ptychographic phaseprojections in atomic electron tomography is presented. Reconstructed phase
images are linear projections of the Coulomb potential and thus are able to
image low-Z atoms at a lower electron dose. Advantages of ptychographic AET are
presented with numerical simulations of the methodology on a 5-nm zinc-oxide
nanoparticle and WS2WSe2 van der Waals heterostructure.
In the field of machine learning, specifically in the field of deeplearning, computer scientists were able to effectively leverage Graphics
Processing Units to train mathematical structures that act as universal function
approximators. These neural networks shocked the world in their power and
versatility: complex image classification, creative text generation, and
calculation of heuristics for games previously thought by mankind as impossible.
This thesis facilitates the merging of Fourier microscopy and machinelearning via the introduction of a novel deep-learning based algorithm that
processes diffraction patterns into meaningful phase image reconstructions.
After a systematic explanation of the algorithm, experimental results from
scanning electron CDI experiments imaging monolayer graphene, twisted bilayer
hexagonal boron nitride, and a gold nanoparticle are presented along with
comparisons via traditional methods.