Data Science for Design of Functional Materials

Wagner, Nicholas Adam

doi:https://doi.org/10.21985/n2-73c2-cx44

Work

Data Science for Design of Functional Materials

Public

Download PDF

Download All Files (.zip)

Functional electronic materials are difficult to design due to the complex interplay among chemistry, atomic structure, and electrical properties. This dilemma is further amplified in transition metal compounds which can defy the band-theory description of non-correlated electrons. Exploring the vast possible design space completely with experiments or first-principles simulations is not practical. Recently, data science has emerged as a possible route to understand and predict properties of materials from known examples. In this thesis, I apply a variety of computational techniques such as density functional theory, data visualization, and machine learning to the design of functional materials. I first show how the change in bond topology of face-sharing perovskite analogues offers new avenues for control of elastic properties and band structure beyond what is achievable in conventional cubic perovskites. I then discuss the limitations of statistical analysis in structure-property studies and how density functional theory simulations can clarify the ways in which datasets are incomplete. Next, I utilize machine learning to construct models of metal-insulator transitions in the periodic table and in a dataset of materials with measured resistivities. I query the models' use of features with SHAP feature importance scores to quantify what the models are learning. Finally, I describe an open-source software package for the generation of training sets for machine-learned interatomic potentials. With these advances, I hope to motivate greater adoption of data-driven methods and open science in the functional materials community.

Creator