cgcnn—a graph representation of materials for property...

14
CGCNN—A Graph Representation of Materials for Property Prediction and Materials Design Tian Xie, Jeffrey Grossman Department of Materials Science and Engineering, Massachusetts Institute of Technology 11/29/2018 1

Upload: others

Post on 21-Apr-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

CGCNN—A Graph Representation of Materials for Property Prediction and

Materials Design

Tian Xie, Jeffrey GrossmanDepartment of Materials Science and Engineering,

Massachusetts Institute of Technology11/29/2018

1

Page 2: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Deep learning for different data types

Grossman’s Group, Massachusetts Institute of Technology

2

Images

Grid-like

Language

Time series

Graph

Relational

CNNTranslation invariance

RNNTime

invariance

GCNNode

invarianceBattaglia, et al., arXiv preprint arXiv:1806.01261 (2018).

Invariances are built into the neural networks to prevent overfitting

Page 3: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Representation of solid materials

Grossman’s Group, Massachusetts Institute of Technology

3

Molecule graphs• Node invariance à atom

indistinguishability • Graph construction is straightforward• All information encoded

Gasteiger, et al. Angew. Chem. Int. Ed. Engl. 32.4 (1993): 503-527.Duvenaud, et al. Adv Neural Inf Process Syst. 2015.Kearnes, et al. J. Comput. Aided Mol. Des. 30.8 (2016): 595-608.Gilmer, et al. arXiv preprint arXiv:1704.01212 (2017).Crystal (solid material) graphs

• 3D structure of materials are important

• Many materials are periodic• Bond determination can be

ambiguous• Rare elements

Innovation needed for graph construction and network design

N

N N

N

O

O

Page 4: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Design of CGCNN

Grossman’s Group, Massachusetts Institute of Technology

4

Tian Xie, and Jeffrey C. Grossman. Phys. Rev. Lett. 120.14 (2018): 145301.

Crystal Graph Convolutional Neural Networks (CGCNN)

¡ 3D structure: use distances as edge attributes

¡ Periodicity: multigraphs instead of normal graphs

¡ Rare elements: element encoding based on elemental properties

¡ Bond ambiguity: ”gated architecture” in GCN

3D molecules:Schütt, et al. Nat. Commun. 8 (2017): 13890.Schütt, et al. J. Chem. Phys. 148.24 (2018): 241722.

Na1 Na1Cl1

Periodic image

Page 5: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Prediction performance

Grossman’s Group, Massachusetts Institute of Technology

5

¡ Materials project: 87 elements, 216 space groups, up to 7 different elements and 200 atoms in unit cell.

¡ Achieve DFT accuracy on test sets for 7 properties with 10# data points.

Tian Xie, and Jeffrey C. Grossman. Phys. Rev. Lett. 120.14 (2018): 145301.

Page 6: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Application: solid state electrolytes

Grossman’s Group, Massachusetts Institute of Technology

6

Zeeshan Ahmad, Tian Xie, et al. ACS Cent. Sci. 4.8 (2018): 996-1006.

¡ Lithium metal batteries¡ Dendrite growth at the interface

between Lithium metal and electrolytes is a big problem

¡ Causes short circuit and capacity loss¡ Design new electrolytes to suppress

dendrite growth

¡ Kinetic model for the stability of the interface¡ Inputs: shear/bulk modulus, molar

volumes¡ Output: Stability parameter !

(negative means stable)

Ahmad, et al. Phys. Rev. Lett. 119.5 (2017): 056003.In collaboration with Zeeshan & Prof. Viswanathan from CMU

Page 7: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Application: solid state electrolytes

Grossman’s Group, Massachusetts Institute of Technology

7

In collaboration with Zeeshan & Prof. Viswanathan from CMU

¡ Screening Li-containing solids¡ Data: ~3400 training data from Materials

Project¡ Predicted stability parameter ! for

12,950 solid materials¡ Identified several candidates

¡ Uncertainty estimation¡ CGCNN ensembles for uncertainty

estimation¡ Extrapolation can cause higher

uncertainties

Zeeshan Ahmad, Tian Xie, et al. ACS Cent. Sci. 4.8 (2018): 996-1006.

Page 8: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Hierarchical visualization of CGCNN

Grossman’s Group, Massachusetts Institute of Technology

8

E0

En

E

¡ Elemental representation¡ Only the information of

elements

¡ Local representation¡ Surrounding elements¡ Local configuration

Tian Xie, and Jeffrey C. Grossman. J. Chem. Phys. 149.17 (2018): 174111.

¡ Local energy¡ Energy of the atom in

local environments

Page 9: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Perovskites: different compositions

Grossman’s Group, Massachusetts Institute of Technology

9

¡ Perovskites¡ A crystal structure

type ABC3

¡ Site A, B: metals¡ Site C: O, F, S, N

Tian Xie, and Jeffrey C. Grossman. J. Chem. Phys. 149.17 (2018): 174111.

Page 10: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Less data needed for element representations

Grossman’s Group, Massachusetts Institute of Technology

10

Tian Xie, and Jeffrey C. Grossman. J. Chem. Phys. 149.17 (2018): 174111.

234 training data (1/64)

937 training data (1/16)

3,750 training data (1/4)

15,000 training data

Page 11: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Boron: different configurations

Grossman’s Group, Massachusetts Institute of Technology

11

Tian Xie, and Jeffrey C. Grossman. J. Chem. Phys. 149.17 (2018): 174111.

Page 12: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Energy of different coordination environments

Grossman’s Group, Massachusetts Institute of Technology

12

Tian Xie, and Jeffrey C. Grossman. J. Chem. Phys. 149.17 (2018): 174111.

Tm-H solid solutions

Special stability due to Al-O, Si-O, P-O, S-O covalent bond

¡ Local energy of all 734,077 distinct coordination environments in Materials Project

¡ General trends¡ Diagonal lines: zero

local energy¡ Most stable: high

valence metals + high electronegativity non-metals à strong ionic bondseV/atom

Page 13: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

Summary

¡ CGCNN provides a general framework to learn structure-property relations for solid materials.

¡ Uncertainty estimation is important when applying ML to real material systems.

¡ Visualization of neural networks can provide additional insights for materials design.

Grossman’s Group, Massachusetts Institute of Technology

13

Page 14: CGCNN—A Graph Representation of Materials for Property ...txie.me/assets/pdfs/talks/2018/MRS-CGCNN.pdfZeeshan Ahmad, Tian Xie, et al.ACS Cent. Sci.4.8 (2018): 996-1006. ¡Lithium

References and codes

¡ References:¡ Xie, Tian, and Jeffrey C. Grossman. "Crystal Graph Convolutional Neural

Networks for an Accurate and Interpretable Prediction of Material Properties." Physical review letters 120.14 (2018): 145301.

¡ Ahmad, Zeeshan, et al. "Machine Learning Enabled Computational Screening of Inorganic Solid Electrolytes for Suppression of Dendrite Formation in Lithium Metal Anodes." ACS central science 4.8 (2018): 996-1006.

¡ Xie, Tian, and Jeffrey C. Grossman. "Hierarchical visualization of materials space with graph convolutional neural networks." The Journal of chemical physics 149.17 (2018): 174111.

¡ Codes:¡ CGCNN: https://github.com/txie-93/cgcnn

Grossman’s Group, Massachusetts Institute of Technology

14

Thank you!