Deep learning methods for the design and understanding of solid materials

Abstract

The trend of open material data and automation in the past decade offers a unique opportunity for data-driven design of novel materials for various applications as well as fundamental scientific understanding, but it also poses a challenge for conventional machine learning approaches based on structure features. In this thesis, I develop a class of deep learning methods that solve various types of learning problems for solid materials, and demonstrate its application to both accelerate material design and understand scientific knowledge. First, I present a neural network architecture to learn the representations of an arbitrary solid material, which encodes several fundamental symmetries for solid materials as inductive biases. Then, I extend the approach to explore four different learning problems: 1) supervised learning to predict material properties from structures; 2) visualization to understand structure-property relations; 3) unsupervised learning to understand atomic scale dynamics from time series trajectories; 4) active learning to explore an unknown material space. In each learning problem, I demonstrate the performance of the approach compared with previous approaches, and apply it to solve several realistic materials design problems and extract scientific insights from data.