Materials discovery using machine learning based on materials database and informatics
File(s)
Author(s)
Chung, Vincent
Type
Thesis
Abstract
The discovery of new materials has been pivotal to technological advancements but is currently bottlenecked by the experimental validation of proposed hypothetical compounds. There are several obstacles to realizing the full potential of data-driven approaches, namely 1) identification of synthesizable materials 2) availability of high-quality data, and 3) omitted practical consideration.
To address these challenges, a dataset on solid-state synthesis of ternary metal oxides was manually curated. This data offers information on whether materials have been solid-state synthesized in the literature, along with some important synthesis conditions. We then applied this dataset in various ways that can help accelerate materials discovery. We first showed that a small set of high-quality data can be used to identify outliers in text-mined datasets. We then trained a transductive positive-unlabeled learning model to predict the solid-state synthesizability of the ternary metal oxides. Compared to previous studies that predict the general synthesizability, the model offers information on false positive rates.
In chapter 4, we trained models using different datasets and features to predict solid-state heating temperature and atmosphere. Comparison of models performance shows that the best overall models were trained using a combination of the two datasets. We also demonstrated that precursor features generated using high-quality data can eliminate the need for precursor information from other sources that are scattered, non-uniform, or incomplete.
In chapter 5, we conducted a screening study that incorporated precursor hazards as a criteria. Precursor hazards are often neglected during materials screening but are important for synthesis planning, lab workers’ safety, and environmental sustainability.
Overall, the work in this thesis provides a set of high-quality datasets to supplement text-mined datasets for solid-state synthesis by illustrating approaches to select hypothetical compositions, offering advice in choosing solid-state synthesis conditions, and raising awareness of practicalities when applying data-driven approaches.
To address these challenges, a dataset on solid-state synthesis of ternary metal oxides was manually curated. This data offers information on whether materials have been solid-state synthesized in the literature, along with some important synthesis conditions. We then applied this dataset in various ways that can help accelerate materials discovery. We first showed that a small set of high-quality data can be used to identify outliers in text-mined datasets. We then trained a transductive positive-unlabeled learning model to predict the solid-state synthesizability of the ternary metal oxides. Compared to previous studies that predict the general synthesizability, the model offers information on false positive rates.
In chapter 4, we trained models using different datasets and features to predict solid-state heating temperature and atmosphere. Comparison of models performance shows that the best overall models were trained using a combination of the two datasets. We also demonstrated that precursor features generated using high-quality data can eliminate the need for precursor information from other sources that are scattered, non-uniform, or incomplete.
In chapter 5, we conducted a screening study that incorporated precursor hazards as a criteria. Precursor hazards are often neglected during materials screening but are important for synthesis planning, lab workers’ safety, and environmental sustainability.
Overall, the work in this thesis provides a set of high-quality datasets to supplement text-mined datasets for solid-state synthesis by illustrating approaches to select hypothetical compositions, offering advice in choosing solid-state synthesis conditions, and raising awareness of practicalities when applying data-driven approaches.
Version
Open Access
Date Issued
2025-02-18
Date Awarded
01/10/2025
License URL
Advisor
Payne, David
Walsh, Aron
Publisher Department
Department of Materials
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)