Projects per year
Abstract
Computational discovery of new materials with specific properties is a significant scientific challenge with societal benefits. This is made difficult by the vast number of possible element combinations in the periodic table and the nearly limitless ways these elements can be arranged in periodic structures. Moreover, mapping structures to properties using computational methods like density functional theory (DFT) is complex and resource-intensive, which makes exploring this material space impractical.
This thesis investigates machine learning techniques to guide the atomistic materials discovery and navigating the complex materials space.
A key achievement of this work is the significant expansion of computationally discovered 2D materials through high-throughput DFT evaluation of a generative model known as CDVAE. This model is benchmarked against traditional lattice decoration methods, with both approaches yielding remarkably similar success rates in producing thermodynamically stable structures, thus establishing CDVAE as an efficient and reliable generative model. As a result of this effort, the number of hypothetical materials in the C2DB is nearly quadrupled by adding 3321 new materials with energies below 0.1 eV/atom of the convex hull.
Furthermore, the dataset is expanded by an additional 918 structures, bringing the total to 4249 novel materials. All of these are characterized using C2DB’s high-throughput workflow for ab initio property characterization.
Additionally, the CDVAE model is used to discover 529 new stable 1D materials, and, thus, doubling the number of stable 1D structures in the C1DB.
The thesis also explores applications of universal interatomic potentials (UIP). The first application integrates UIPs with Bayesian active learning in the BEACON framework to find global energy minima of atomic structures. Notably, one UIP demonstrates significant performance improvements across various systems. The second application uses a UIP to find the optimal adsorption configuration of a terrylene on 2D materials, which allows for the study the photoemission spectrum, which is in good agreement with experimental data.
Finally, a machine learning classifier is developed to predict dynamical stability using electronic descriptors, achieving an area under the curve score of 0.90, making it suitable as a screening step in high-throughput workflows.
This thesis investigates machine learning techniques to guide the atomistic materials discovery and navigating the complex materials space.
A key achievement of this work is the significant expansion of computationally discovered 2D materials through high-throughput DFT evaluation of a generative model known as CDVAE. This model is benchmarked against traditional lattice decoration methods, with both approaches yielding remarkably similar success rates in producing thermodynamically stable structures, thus establishing CDVAE as an efficient and reliable generative model. As a result of this effort, the number of hypothetical materials in the C2DB is nearly quadrupled by adding 3321 new materials with energies below 0.1 eV/atom of the convex hull.
Furthermore, the dataset is expanded by an additional 918 structures, bringing the total to 4249 novel materials. All of these are characterized using C2DB’s high-throughput workflow for ab initio property characterization.
Additionally, the CDVAE model is used to discover 529 new stable 1D materials, and, thus, doubling the number of stable 1D structures in the C1DB.
The thesis also explores applications of universal interatomic potentials (UIP). The first application integrates UIPs with Bayesian active learning in the BEACON framework to find global energy minima of atomic structures. Notably, one UIP demonstrates significant performance improvements across various systems. The second application uses a UIP to find the optimal adsorption configuration of a terrylene on 2D materials, which allows for the study the photoemission spectrum, which is in good agreement with experimental data.
Finally, a machine learning classifier is developed to predict dynamical stability using electronic descriptors, achieving an area under the curve score of 0.90, making it suitable as a screening step in high-throughput workflows.
Original language | English |
---|
Publisher | Department of Physics, Technical University of Denmark |
---|---|
Number of pages | 178 |
Publication status | Published - 2024 |
Fingerprint
Dive into the research topics of 'Machine learning for atomistic materials discovery'. Together they form a unique fingerprint.Projects
- 1 Finished
-
Machine learning for atomistic materials discovery
Lyngby, P. M. (PhD Student), Thygesen, K. S. (Main Supervisor), Jacobsen, K. W. (Supervisor), Botti, S. (Examiner) & Kermode, J. (Examiner)
01/09/2021 → 14/01/2025
Project: PhD