Studying complex microstructures with advanced transmission electron microscopy and machine learning

Abstract

This thesis demonstrates the development of unsupervised machine learning workflows on correlated advanced scanning transmission electron microscopy (STEM) datasets for characterization of nano-scale microstructure found across a wide range of material systems. A key goal in materials science is to link a material’s physical characteristics to its properties. Recent advancements in TEM have made it possible to rapidly and simultaneously capture multiple signals generated by electron interactions with the specimen. Techniques like 4D-STEM and STEM-EDS now enable the collection of structural and compositional data at nanoscale to atomic resolution across the sample. In 4D-STEM, diffraction patterns are recorded at each scan point, while STEM-EDS gathers elemental spectra in a similar manner. These scans typically include tens of thousands of points, making traditional analysis methods challenging, even for experienced researchers. Often, datasets are visually mapped and qualitatively compared, potentially overlooking important quantitative relationships. Additionally, users may underutilize the full range of available data streams for the sake of simplicity. Such approaches have limitations, as leveraging information across multiple dimensions with appropriate quantitative methods could help uncover statistically significant features in otherwise noisy datasets. The first part of the project showcases the initial attempt at developing an unsupervised clustering workflow that identifies quantitative correlations between simultaneously acquired 4D-STEM and STEM-EDS datasets of a Co2FeSi Heusler alloy system. The combination of both managed to identify a secondary phase coherent precipitate that could not be determined from either the individual EDS or diffraction data alone. The second part of the project involves a comparison between the merged correlative unsupervised clustering workflow, with another unsupervised machine learning workflow that worked with 4D-STEM in its image form, which was shown to have worked in previous studies. The two workflows were used to analyse the Co2FeSi mentioned earlier, as well as a simulated Au−Al2Cu matrix-precipitate system. In both cases, the stoichiometry differences are difficult to segment using EDS. For the former system, the same elements are present in both the primary and secondary phase, while for the latter system, the relatively lighter Al2Cu precipitates will produce weak EDX signal compared to the heavy Au matrix. In terms of diffraction, both systems have common reflections between different phases, making it difficult to partition. The two workflows were evaluated using a set of unsupervised clustering metrics, and it was determined that the merged correlative unsupervised clustering workflow produced better results in identifying small but structurally significant microstructure features. The third part of the project involves an investigation into the different types of 4D-STEM feature engineering techniques on a Scanning Precession Electron Diffraction dataset of TRistructural ISOtropic particles. It was found that there was not a universal feature set that could characterize all important microstructure features present in the dataset, which motivated the development of a novel consensus clustering approach that attempts to aggregate different cluster results produced from the different feature sets. The last part of the project demonstrates the feasibility of automating high-magnification acquisition of simultaneous 4D-STEM and STEM-EDS datasets on bio-synthesized Pd-Cu nanoparticles identified by Convolutional Neural Network from a relatively low-magnification High Angular Annular Dark Field STEM image. The acquired datasets are then analyzed using the merged correlative unsupervised clustering workflow to reveal nanoscale variations within each nanoparticle.

Date of Award	19 Jun 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Alexander Eggeman (Main Supervisor) & Ping Xiao (Co Supervisor)