Data-driven discovery of low-dimensional materials and crystal structure search
Abstract/Contents
- Abstract
- Materials are at the heart of the technological challenges faced by humanity --technologies such as quantum computing, energy storage rely on specific functionalities of unique materials. With the rise of large-scale computations in materials science in the past decade, scientists are now able to discover and design materials at an unprecedented pace through data-driven approaches. With methods such as high-throughput computations, data mining and machine learning, the field of computational materials science is undergoing a paradigm shift in the way we model new materials. We explore three examples of such data-driven discovery of new materials, starting with new low-dimensional materials. Two-dimensional materials have garnered a wide range of research interest in the past decade. Their unique structure, composed of strongly bonded sheets of atoms held together by weak van der Waals forces, allows for both technological applications and gives rise to physics phenomena observed only in two dimensions. But only a few dozen layered materials have been identified as such, though there are likely to be many more that could have superior properties. This leads to a bottleneck in the research of two-dimensional materials. First, we identify a broad spectrum of two-dimensional layered materials using a novel data mining algorithm that screens for two-dimensional materials from databases of crystal structures. The algorithm uses interatomic distances as a heuristic for atomic bonds and identifies weakly bonded subunits in the crystal. Then the dimensionality of the weakly bonded subunits is determined. We use this algorithm to screen crystals in the Materials Project database and identify 1173 two-dimensional layered materials, 487 one-dimensional materials and 98 bulk intrinsic heterostructures. This work substantially increases the scope of two- and one-dimensional materials, and we provide material properties including band gaps and point groups to facilitate screening for desired applications. Data mining is limited to materials that have already been discovered and put in databases and is subject to bias in the data. In the second part, we extend the search for two-dimensional materials to materials that have yet to be discovered or synthesized. We use machine learning to predict chemical composition of two-dimensional materials and screen billions of candidates to discover the full spectrum of possible two-dimensional materials. Out of over 1000 newly identified two-dimensional materials, we verify a small number of candidates using density functional theory including two phase-change material candidates. Our model performs five times better than scientists working in the field and is comparable to best solid-state chemists. We achieve this performance using semi-supervised learning, as materials with the same chemical formula may have multiple phases giving rise to ambiguity in labels. We demonstrate that semi-supervised learning may be a useful tool for material science where labeled data is often scarce. In the last part, we note that many machine-learned predictions of new materials, including our model for two-dimensional materials, predict chemical compositions of new materials but precludes structure. Though crystal structures determine the physical and chemical properties of materials, finding energetically stable crystal structures of a material given the chemical formula remains an unsolved challenge in materials science. State-of-the-art methods for crystal structure search involve thousands of computationally expensive density functional theory calculations, making it intractable for most materials design problems. We accelerate crystal structure search by learning to relax randomly sampled crystal structures using graph neural networks. We generate a high-throughput density functional theory dataset of more than 100,000 random structure relaxations and train a model that calculates forces and stresses on crystals on the structures seen during the relaxations. We show that models trained on data conventionally used to train interatomic potentials fail to simulate relaxations from random structures and using random structure relaxations data leads to up to two orders of magnitude decrease in mean absolute error on forces and stresses for the same task.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2021; ©2021 |
| Publication date | 2021; 2021 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Cheon, Gowoon |
|---|---|
| Degree supervisor | Heinz, Tony F |
| Degree supervisor | Reed, Evan J |
| Thesis advisor | Heinz, Tony F |
| Thesis advisor | Reed, Evan J |
| Thesis advisor | Devereaux, Thomas Peter, 1964- |
| Degree committee member | Devereaux, Thomas Peter, 1964- |
| Associated with | Stanford University, Department of Applied Physics |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Gowoon Cheon. |
|---|---|
| Note | Submitted to the Department of Applied Physics. |
| Thesis | Thesis Ph.D. Stanford University 2021. |
| Location | https://purl.stanford.edu/cg641qk8811 |
Access conditions
- Copyright
- © 2021 by Gowoon Cheon
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...