Electron Microscopy (EM) has become a tool of choice to quantify materials structure at high resolution. Four-Dimensional Scanning Transmission Electron Microscopy (4D-STEM) is a versatile diffraction EM technique that can provide a high spatial resolution map of material structures and properties. However, 4D-STEM datasets can contain hundreds of thousands of patterns and can thus be challenging to manually interpret. Robust computational tools are required to gain insights from 4D-STEM datasets.

In this work, we present a set of computational methods for interpreting 4D-STEM datasets. There are two broad categories these methods fall into. The first category, unsupervised learning, allows for similarities across a dataset to be mapped with no prior knowledge regarding the structures in the dataset. To perform unsupervised learning on 4D-STEM datasets, we implement feature engineering protocols that distill the information in each individual diffraction pattern down. We then analyze the performance of both Non-negative Matrix Factorization (NNMF) and consensus clustering on a set of simulated datasets to highlight the versatility of different feature extraction techniques.

The second category of computational methods presented in this work is based on template-matching procedures. If the crystal structure(s) in a material are known, templates can be used to quantify the structures present in each diffraction pattern. Here, we present preliminary work towards mapping the spatial distribution of phases in multi-phase materials using Non-negative Least Squares (NNLS). Together, unsupervised and template-matching approaches allow researchers using 4D-STEM to extract information from their datasets more readily.

Precisely controlled synthesis of nanostructures is heavily emphasized in the field of nanoscience, in large part due to the desire to control the size, shape, and terminating facets of nanoparticles for applications in catalysis, optics, and medicine. Direct control of the size and shape of solution-grown nanoparticles relies on an understanding of how synthetic parameters alter nanoparticle structures during synthesis. Synthesis conditions rarely yield uniform particles, but the heterogeneity in these populations is hard to quantify especially in respect to atomic structure. Defects are a particularly challenging element to quantify the influence of synthesis on due to the limited number of methods which can give insight into these features. In this thesis we analyze the development of defect structures in multiply twinned metal nanoparticles. We find that non-classical growth mechanisms occur during standard colloidal growth. In order to further study the conditions which cause these non-classical growth modes we analyze methods to automate the analysis of defects in nanoparticles. To this end we propose a machine learning based pipeline for the segmentation and stacking fault classification in nanoparticles. We demonstrate a flexible two step pipeline for analysis of high resolution transmission electron microscopy data, which uses a U-Net for segmentation followed by a random forest for detection of stacking faults. Our trained U-Net is able to segment nanoparticle regions from amorphous background with a Dice coefficient of 0.8 and significantly outperforms traditional image segmentation methods. Using these segmented regions, we are then able to classify whether nanoparticles contain a visible stacking fault with 86% accuracy. We further explore the role of architecture in achieving better segmentation of HRTEM micrographs. These studies in turn illuminate the challenges in generating a large enough dataset for neural network training in electron microscopy. To solve this we evaluate methods to minimize labelling requirements by using automated labeling methods. This strategy recasts the classification problem as a label noise problem, with the fraction of label noise problem. We analyze the impact of label noise rate and training method on the classification error of neural networks for SEM data. We compare the noise resistance of three training methods: standard network training, pretraining with a tiny, label noise-free dataset, and co-teaching. We find that the pretraining approach yields the most accurate results across label error rates. These developments will help in enabling automated analysis of nanomaterials.