Files in this item



application/pdfSCHILLER-DISSERTATION-2020.pdf (11MB)
(no description provided)PDF


Title:Data-driven materials design
Author(s):Schiller, Joshua
Director of Research:Ertekin, Elif
Doctoral Committee Chair(s):Ertekin, Elif
Doctoral Committee Member(s):Johnson, Harley; Shoemaker, Daniel; Tawfick, Sameh
Department / Program:Mechanical Sci & Engineering
Discipline:Mechanical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
machine learning
neural network
materials discovery
Abstract:In an era of increasing global temperatures and an ever-growing desire for higher performance technology, the need for superior materials and accelerated manufacturing processes has never been greater. The production of better solar power generation will require new photovoltaic materials. Advanced devices and sensors will require new electronic materials. And in a globalized economy, the pressures to hasten the discovery process are only growing. These factors have led to an increasing focus on inverse design. Unlike in traditional approaches, where a material is first discovered and then an application is found, the goal of inverse design is to instead generate an optimal material for a desired application — even if the material is not previously known. While progress has been made, there is still a long road to an ideal inverse design process. For instance, the ability to quickly and accurately design new crystalline materials remains a longstanding challenge in the materials research community. Traditionally, many materials have been discovered by accident or with experimental trial-and-error. This process can often be unreliable, labor-intensive and/or cost-intensive. Some scientists have turned towards theoretical methods that utilize physical simulations to augment the search process. These techniques require high throughput calculation of numerous potential compounds, necessitating excessive computational resources. There have been improvements on these methods that have led to algorithms to try to minimize the domain of potential stable compounds. However, many of these methods still require large numbers of calculations and struggle with more complex systems. Moreover, the physical simulations themselves that these algorithms use can introduce a systematic error that impedes the ability to accurately predict stability. To address these challenges, in this thesis I demonstrate several approaches to data-driven materials and manufacturing, focusing on computational tools, digital data, and experimental tools. In the space of computational tools, an approach that combines quantum mechanical modeling and machine learning is demonstrated as a technique for the accelerated discovery of spinel compounds. In the space of digital data, I introduce a platform to collect, analyze and learn from graphene synthesis data collected from the research community. For experimental tools, I construct an artificial neural network model to automatically detect regions of electron microscopy images that contain graphene. Computational tools. I explore a data-driven materials design as an alternative technique for discovering new spinel compounds. By combining data from the literature with the physical simulations, I mitigate the error introduced by the simulations themselves. I also utilize reinforcement learning in the form of an iterative procedure. The process simulates top candidates and updates the model with new information gleaned from the energy calculations. I show that the resulting model demonstrates a comparable performance to a pure simulation approach at a fraction of the computational cost. Due to the iterative nature of the process that only selectively simulates a subset of all candidate structures, comparatively fewer calculations are used than a brute-force approach. Digital data. In the latter parts of my thesis, I investigate data-driven design as applied to graphene synthesis. Since its discovery in 2004, graphene has captivated the research community. However, despite its potential, the manufacturing of pristine, high quality graphene has still not reached the point of commercial viability. This is in part due to the nature of graphene synthesis. Graphene is typically synthesized using chemical vapor deposition (CVD), a procedure whereby carbon-containing precursors are fed into a vacuum furnace and react with a catalyst under a particular set of environmental conditions to form the graphene. This growth process is notoriously sensitive, with small deviations in the procedure entailing drastically different results. Moreover, the CVD procedures that have been tried and their resulting products are distributed throughout the literature, which makes it difficult to detect patterns from prior work. The information that is provided is also not standardized, which makes direct comparisons between synthesis recipes even more challenging. Furthermore, many graphene recipe results are not reported at all, particularly if the results are not satisfactory. To address this challenge, I introduce the “Graphene — Recipes for the synthesis of high-quality material” (Gr-ResQ: pronounced graphene rescue), a platform providing a central repository for graphene synthesis data as well as tools to aid the researchers themselves. Gr-ResQ crowd-sources graphene recipe data and associated results, providing a means for researchers to learn from the collective knowledge of the research community. Moreover, it provides tools to more easily post-process their SEM and Raman spectroscopy data. The ultimate goal of Gr-ResQ is to build a decentralized sequential learning procedure, akin to the aforementioned spinel work. Experimental tools. To learn from the graphene data still requires an objective function to measure the success of a recipe. As one of the key metrics of a successful recipe is the quantity of graphene produced, acquiring the surface area of the synthesized graphene is crucial to any future predictive model. As surface area is typically extracted from SEM data, it is necessary to determine regions of an image that contain graphene as opposed to substrate. However, this analysis can be a time-consuming process, which makes post-processing difficult to conduct at scale. To expedite this, I have also developed a neural network to automatically classify regions of SEM images that contain graphene. I utilize a U-Net encoder/decoder architecture that is trained on data collected using Gr-ResQ's image tool. The fitted model is then shown to provide very high accuracy results. With such performance, the model can eventually be incorporated directly into Gr-ResQ to automatically post-process SEM images that are ingested into the database.
Issue Date:2020-08-06
Rights Information:Copyright 2020 Joshua Schiller
Date Available in IDEALS:2021-03-05
Date Deposited:2020-12

This item appears in the following Collection(s)

Item Statistics