UC Berkeley

Electron microscopy is an important technique for the characterization of nanostructures and materials. Since the rise of nanosciences, electron microscopy has also experienced significant rise in its importance, as it is the only characterization technique that allows one to directly visualize the unique nanoscale structures of nanomaterials. Recent advancement has allowed electron microscopic data acquisition to be automated, significantly increasing the rate in which these data are produced. Manual analysis of electron microscopic data is a time consuming process, and is insufficient in the face of an automated electron microscope. An efficient electron microscopic data analysis algorithm not only can allow physically relevant information to be obtained from electron microscopy at an unprecedented rate, but also could lead to a fully autonomous electron microscopy system, in which data acquisition, data analysis, and the design of new experiments can be done on-the-fly without the need of human intervention. A fully autonomous electron microscopy system has the potential to revolutionize research in material sciences.Machine learning provides a solution to the challenge of data analysis for autonomous electron microscopy. Unsupervised machine learning can extract physically relevant information from complex electron microscopic data with minimum requirement on human input and intervention. As such, unsupervised machine learning algorithms can potentially serve as the on-the-fly decision making agents in an autonomous electron microscopy system. This thesis introduces two unsupervised learning algorithms that can perform on-the-fly decision making for an autonomous electron microscope. In this thesis, Chapter 1 outlines the state-of-art of automated electron microscopy and existing machine learning algorithms for electron microscopic data analysis. This chapter will also discuss the demands for machine learning in autonomous electron microscopy, and what kind of machine learning algorithms are needed to meet such demands. Chapter 2 introduces AutoDetect-mNP, an unsupervised machine learning algorithm capable of extracting and classifying the shapes of nanoparticles from electron microscopic images. AutoDetectmNP can serve as an on-the-fly decision making agent when the autonomous electron microscope is screening a sample at low magnification. Chapter 2 introduces a variational autoencoders model for information extraction from atomic resolution in situ electron microscopy data. Variational autoencoders can serve as an on-the-fly decision making agent when an autonomous electron microscope is studying nanoscale dynamics at atomic resolution. Finally, Chapter 4 covers the outlooks for autonomous electron microscopy. Chapter 4 proposes several potential future directions an autonomous electron microscopy system can be applied in. Outlooks of these research directions are discussed and discussed are also how autonomous electron microscopy will impact the field of material sciences as a whole.