ABSTRACT
Scanning (transmission) electron microscopies (S(T)EM) and spectroscopies are a well-established, robust, set of imaging tools that proved to be irreplaceable in visualizing structure and mapping chemical identity at atomic resolution. The goal of these highly localized imaging and spectroscopy experiments is to observe and correlate structure-property relationships with functionality by evaluating chemical, electronic, optical, and phononic properties of individual atoms as well as nanoscale structural elements. Improvements in the instrument hardware and data processing has allowed many of these properties to be mapped with sub-10-pm precision, which enabled the visualization of chemical and mechanical strains, as well as parameters such as ferroelectric polarization and octahedral tilts. However, these improvements come at a cost, with the negative side effects rooted in data variability, velocity, and volume. This wealth of extracted information necessitates a drastic improvement in the capability to transfer, store, and analyze multidimensional datasets. Here, data interpretation is perhaps the biggest challenge because the other factors are mainly engineering challenges with existing (albeit currently still expensive) solutions. In this chapter, we explore the technologies, methods, and algorithms capable of tackling these tasks as well as discuss approaches that may significantly reduce the data, the analysis, and the interpretation bottlenecks. Specifically, we will focus on supervised and unsupervised machine learning, along with a brief discussion on broader impacts, including open data formats, community-driven software, and data repositories.
