Data mining for materials discovery is concerned with representing materials science problems into a statistical framework, and learning models that describe observations about the processing, structure, and property of materials. The type of materials includes metals, ceramics, glass, polymers, and composites which are mixtures of multiple types. Observations come from either computational simulation or laboratory experiments. The aim is to analyze the observational datasets to find relationships, and to present them in ways that are both understandable and useful. The quality of data plays an important role in data mining practice; there can be multiple sources of signals creating multiple contexts in data. This Ph.D. thesis outlines the problem of building both the representation, and the core of learning in the process of materials design and discovery, from a rather general, agnostic point of view by use of data mining strategies. A particular emphasis is on how to detect and model complex contextual structures in data. We start with an optimization problem, as optimization is the core to any machine learning algorithm. We present a learning system that helps solve optimization problems faster, with techniques like supervised region reduction and feature ranking. Then we study the problem of finding a better representation method for designing heterogeneous microstructures. Next, we explore supervised learning to construct models that predict lower level response from higher level structure. Further explorations feature the application of deep neural networks in both representation and modeling phases of materials systems.

Last modified

• 02/13/2018

Creator

• Ruoqian Liu

DOI

• https://doi.org/10.21985/N2809D

Subject

• Computer Engineering

Keyword

• Material Science
• Applications
• Computer Engineering

Date created

• 2016-01-01

Resource type

• Dissertation

Rights statement

• In Copyright