Development and Expansion of Tools for Data-Driven Materials Development

Abstract

Machine learning and natural language processing techniques are being integrated into chemistry and materials science, finding utility at field and domain levels of research. While these tools have existed, the relative recent emergence of these tools within high-level programming languages like Python means that they have only recently begun to be utilized at scale. In this dissertation, I explore the ways in which these tools can be applied in field-specific settings and a general, domain-level one. In one, I develop a new analysis methodology utilizing image registration, dimensionality reduction, and multivariate analysis to derive information from multimodal atomic force microscopy images. In a second, I utilize and develop reusable code for a Python package within the scanning probe community to obtain insights about and examine impacts of different physical contributions to a measured signal in a specialized atomic force microscopy technique. In another, I introduce a practitioner-centric framework for evaluating topic models that moves away from the dichotomic approach utilized in model development with a critical downstream benefit of advancing data-driven materials research via natural language processing. These works illustrate the ways in which existing machine learning and natural language processing are powerful tools and makes a case for the need of domain expertise in their development, much like the symbiotic work of computationalists, experimentalists, and theorists.