About this Abstract |
Meeting |
2020 TMS Annual Meeting & Exhibition
|
Symposium
|
Computational Discovery and Design of Emerging Materials
|
Presentation Title |
Predicting the Properties of Crystals with High Accuracy Using Deep Learning |
Author(s) |
Weike Ye, Chi Chen, Zhenbin Wang, Iek-Heng Chu, Yunxing Zuo, Chen Zheng, Shyue Ping Ong |
On-Site Speaker (Planned) |
Weike Ye |
Abstract Scope |
Being able to predict the properties of crystals is a prerequisite to materials discovery. First-principle approaches, such as density functional theory (DFT), are expensive and scale poorly with system size. Here, we demonstrate that machine learning (ML) can provide a shortcut. We show that neural networks (NN) with only two descriptors, the electronegativity and ionic radius, can predict crystal stability within DFT error. The mean absolute errors of the NN-predicted DFT formation energies for two sample system types, garnets and perovskites, are only 7-13 meV/atom and 30-34 meV/atom, respectively. We also developed universal MatErials Graph Network (MEGNet) models that can predict properties for both molecules and crystals. MEGNet models provide highest accuracy to-date in predicting 11 of 13 properties of the molecules in QM9, and 4 properties of crystals in Materials Project. Furthermore, we provide successful approaches to often-encountered problems in materials informatics such as disordered structures and data limitation. |
Proceedings Inclusion? |
Planned: Supplemental Proceedings volume |