Design For Accessibility: Getting it right from the start
NIST-JARVIS infrastructure for Improved Materials Design
1. NIST-JARVIS infrastructure for Improved
Materials Design
Kamal Choudhary
https://jarvis.nist.gov/
NIST, Gaithersburg, MD, USA
CECAM workshop, 10/11/2022
1
Joint Automated Repository for Various Integrated Simulations
2. Acknowledgement and Collaboration
2
A. Biacchi
(NIST)
D. Wines
(NIST)
R. Gurunathan
(NIST)
B. DeCost
(NIST)
Bobby sumpter
(ORNL)
A. Agarwal
(Northwestern
University)
S. Kalidindi
(GAtech)
A. Reid
(NIST)
Ruth Pachter
(AFRL)
Karen Sauer
(George
Mason University)
K. Garrity
(NIST)
David Vanderbilt
(Rutgers
University)
Sergei
Kalinin
(ORNL)
F. Tavazza
(NIST)
8. JARVIS-DFT: Electronic structure calculations
• Schrödinger equation for electrons: wave–particle duality,
• Schrödinger equation of a fictitious system (the "Kohn–Sham system") of non-interacting
particles (typically electrons) that generate the same density as any given system of interacting
particles
• Uses density vs wavefunction quantity
• Although a complete theory, several approximations such as:
1) K-points, 2) vdW interactions, 3) kinetic energy deriv., 4) spin-orbit coupling, 5) e-ph coupling
(Convergence, OptB88vdW, TBmBJ, SOC topology, Superconducting prop. )
r
r
E
r
r
V
m
i
i
i
Eff
2
2
2
XC
ee
Ne
Eff V
V
V
T
V
E
H
Walter Kohn (2013)
Exchange-correlation
8
Many DFT databases with GGA-PBE, fixed k-point, no-SOC, …
9. JARVIS-DFT
9
Motivation: Functional and structural materials design using quantum mechanical methods
~70000 materials, millions of calculated properties, compared with experiments if possible
https://jarvis.nist.gov/jarvisdft/
10. JARVIS-DFT MatProj. OQMD
#Materials (Struct., Ef, Eg ) 70870 144595 (41697 common) 1022663
DFT functional/methods vdW-DFT-OptB88, TBmBJ, DFT+SOC GGA-PBE, PBE+U, GLLBSC GGA-PBE, PBE+U
K-point/cut-off Converged for each material Fixed (1000-3000) kp/atom, 520 eV Fixed kp/atom, cutoff
SCF convergence criteria Energy & Forces Energy Energy
Elastic tensors & point phonons 17402 14072 -
Piezoelectric, IR spectra 4801 3402 -
Dielectric tensors (w/o ion) 4801 (15860) 3402 -
Electric field gradients 11865 - -
XANES spectra - 22000 -
2D monolayers 1011 - -
Raman spectra 400 50 -
Seebeck, Power F 23210 48000 -
Solar SLME 8614 - -
Spin-orbit Coupling Spillage 11383 - -
WannierTB 1771 - -
STM images 1432 - -
11. K-point convergence
11
• Energy per cell convergence of 0.001 eV/cell for each material
• Most DFT high-throughput workflows use per reciprocal atom (pra) =>1000
12. vdW interactions: 3D, 2D, 1D & 0D materials
• vdW materials: high lattice error, is converse true?
• Van der Waals (vdW) bonding in x, y, z-directions; exfoliation energy
• If the error => 5%, we predict them to be low-D materials,
• 1100 mats. with OptB88vdW functionals, tight DFT convergence
• Improved lattice parameters with OptB88vdW
ICSD
ICSD
PBE
l
l
l
12
3D: Si 2D: MoS2
0D: BiI3
1D-MoBr3
Nature: Scientific Reports, 7, 5179 (2017)
Nature:Scientific Data 5, 180082 (2018)
Phys. Rev. B, 98, 014107 (2018)
13. MetaGGA & optoelectronic properties
13
• Bandgap, frequency dependent dielectric function from OptB88vdW (OPT) and Modified Becke-Johnson formalisms (MBJ)
• MBJ gives excellent bandgap with low computational cost, also better dielectric function with linear optics
Nature:Scientific Data 5, 180082 (2018)
~20000 TBmBJ bandgaps and dielectric function
MAE bandgap (eV):
• MatProj: 1.45
• AFLOW: 1.23
• OQMD: 1.14
• OptB88vdW: 1.33
• TBmBJ: 0.51
• HSE06: 0.41
(wrt 54 exp. data)
14. Solar cells & linear optics
14
Scientific Data 5, 180082 (2018)
Chemistry of Materials, 31, 15, 5900 (2019).
Spectroscopic Limited Maximum Efficiency (SLME)
15. Spin-orbit coupling & Topological Materials
New class of materials
(electronic bandgap perspective)
15
Email: kamal.choudhary@nist.gov
https://phys.org/news/2014-01-quantum-natural-3d-counterpart-graphene.html
https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcSzMKD5ICIkR9neJRre3prqIjp_iqLMu6TQp7mXKJqmmh-HqjFB
(2016 Nobel prize)
Metal
Semiconductor
Insulator
16. Spin-orbit Spillage
• Majority of the topological materials driven by spin-orbit coupling (SOC)
• Simple idea: Compare wavefunctions of a material with and without SOC?
• Spillage initially proposed for insulators only, now extended to metals also
• Advantages over symmetry-based approaches:
disordered and magnetic mats.
• For trivial materials, spillage 0.0, non-trivial materials ≥ 0.25
16
https://www.ctcms.nist.gov/~knc6/jsmol/JVASP-1067
𝜂 𝐤 = 𝑛𝑜𝑐𝑐(𝐤) − Tr 𝑃 ෨
𝑃 ; 𝑃 𝐤 =
𝑛=1
)
𝑛𝑜𝑐𝑐(𝐤
ۧ
|𝜓𝑛𝐤 ൻ𝜓𝑛𝐤|
Sci. Rep., 9, 8534 (2019)
NPJ Comp. Mat., 6, 49 (2020)
Phys Rev B, 103, 054602 (2021)
18. 18
BCS Superconductors & E-Ph coupling
Debye
Temp
DOS at
EFermi
https://arxiv.org/abs/2205.00060
Superconductors: Materials to conduct electricity without energy loss when they are cooled below a critical temperature, Tc
MgB2 (Tc = 39 K): Highest Tc ambient condition conventional superconductor
19. Other electronic structure databases
19
JARVIS-TB3Py
Tight-binding models
JARVIS-QMC
Quantum Monte Carlo
22. 22
From ANNs to Graph Convolution Networks
𝑧[𝑙]
= 𝑊
[𝑙]
𝑎
[𝑙−1]
+ 𝑏
[𝑙]
𝑎[1]
= σ( 𝑧[𝑙]
); 𝑎[0]
= 𝑋
1) Forward propagation
2) 𝐶𝑜𝑠𝑡, 𝐽(𝑊, 𝑏) = 𝑓(𝑦 −
𝑦)
X1
X2
Hidden
Layer
Input
Layer
Output
layer
𝑦
3) Gradient descent (∇J):
minimize cost with W,b
4) Backpropagation:
chain rule to get,
𝜕𝐽
𝜕𝑊
1) Convolution:
element-wise multiplication & sum
2) Pool: Max, Average, Sum
3) Fully Connected: Standard NN
Shared weights (Learnable filters),
regularized version of NNs
መ
𝐴 = ෩
𝐷−
1
2 𝐴෩
𝐷−
1
2 𝐻𝑙+1 = σ 𝑊 መ
𝐴𝐻𝑙
Adjacency matrix, A
1 0 1
0 1 0
1 0 1
1) Adjacency matrix, N x N (N: #nodes),
2) D: degree of node
3) Update node representation using
message passing, GPU efficient
4) Update equation is local, neighborhood
of a node only, independent of graph size
Standard NN ConvolutionNN GraphConvNN
Types: un/weighted, un/directed, line,
Hetero/Homogenous, Multigraph
23. 23
Line Graph
Explicitly represent pairwise and triplet (bond angle) interactions using line graph
Possible to extend for n-body, e.g. line graph of line graph
nisaba.nist.gov Tesla V100
25. 25
Performance on the Materials Project Dataset
Trained on 69239 materials (DFT data)
#Epochs: 300
Batch_size: 64
• ~44 % improvement by ALIGNN with similar/better training speed
• Similar performance enhancement on QM9 molecule dataset
• Also available on MatBench: https://matbench.materialsproject.org
26. 26
Performance on the JARVIS-DFT Dataset
Trained on ~55k materials
Total energy, Formation energy , Ehull
Bandgap (OPT), Bandgap (MBJ)
Kv, Gv
Mag. mom
єx (OPT/MBJ), єy (OPT), єz (OPT), є
(DFPT:elec+ionic)
Max. piezo. stress coeff (eij)
Solar-SLME (%)
Topological-Spillage
2D-Exfo. energy
Kpoint-length
Plane-wave cutoff
Max. Electric field gradient
avg. me, avg. mh
n-Seebeck, n-PF, p-Seebeck, p-PF
27. 27
Evac with ALIGNN Energy model
No ML training defect structures/data ! Directly predicting with energy/atom model
Total 508 datapoints, MAE wrt Exp. for subset: 0.3 eV
(Elemental solids+Alloys+Oxides+2D monolayers)
~34 % improvement with scissor shift
https://arxiv.org/abs/2205.08366
pretrained.py --model_name jv_optb88vdw_total_energy_alignn--file_format poscar --file_path POSCAR
28. 28
BCS Superconductors
• Prediction on 10 % test data
• 8293 out of 431778 materials in COD as superconductors
• First predicting Eliashberg function, then Tc 6 % improvement
• ALIGNN for both scalar and spectral learning
Best
32. 32
CO2 Isotherms: AI for Climate Change
DL model for predicting CO2 adsorption in MOFs (using hMOF GCMC data)
Choudhary et al., Computational Materials Science 210, 111388 (2022)
35. 35
Scanning Transmission Electron Microscope Image
PPdSe: JVASP-6316
C: JVASP-667 FeTe: JVASP-6667
Convolution approximation: accurate for thin films mainly (here 2D mats.)
Based on Rutherford scattering model
36. 36
Image classification and semantic segmentation
2D Bravais lattice classification (DenseNet):
1) hexagonal, 2) square, 3) rectangle, 4) rhombus, 5) parallelogram
Baseline accuracy 1/5 = 20 %
Semantic segmentation using U-Net:
Atom vs background, pixelwise classification
38. Background: Feynman’s seminal papers
38
http://physics.whu.edu.cn/dfiles/wenjian/1_00_QIC_Feynman
“Nature is quantum, goddamn it! So if we
want to simulate it, we need a quantum
computer.”
39. Variational Quantum Eigensolver (VQE) &
Variation Quantum Deflation(VQD)
39
http://openqemist.1qbit.com/docs/vqe_microsoft_qsharp.html
Notes:
• Quantum computers are good in preparing states, not good at sum, optimizers, multiplying etc.
• QC to prepare a wavefunction ansatz of the system and estimate the expectation value
VQD: Deflate other eigensatets once ground state is found using VQE
VQE: a hybrid classical-quantum algorithm using Ritz variational principle
40. Typical Flowchart
40
https://github.com/usnistgov/jarvis
https://github.com/usnistgov/atomqc
K. Choudhary, J. Phys.: Condens. Matter 33 (2021) 385501
Wannier functions:
• Complete orthonormalized basis set,
• Acts as a bridge between a delocalized plane wave representation and a localized atomic orbital basis
• All major density functional theory (DFT) codes support generation WFs for a material
𝐻 = ℎ𝑃𝑃
𝑃∈ 𝐼,𝑋,𝑌,𝑍 ⨂𝑛
𝐻𝑗 = 𝐻 + 𝛽𝑖|𝜓(𝜽0
∗)ۧ 𝜓(𝜽0
∗)|
𝑗−1
𝑖=0
𝐺(𝑘, ꞷ𝑛) = [ꞷ𝑛 + 𝜇 − 𝐻(𝑘) − 𝛴(ꞷ𝑛)]−1
http://www.wannier.org/
42. FCC Aluminum Example
42
a) Monitoring VQE optimization progress with several local optimizers such COBYLA, L_BFGS_B, SLSQP, CG, and SPSA
for Al electronic WTBH and at X-point.
b) Electronic bandstructure calculated from classical diagonalization (Numpy-based exact solution) and VQD algorithm for
Al.
c) Phonon bandstructure for Al
43. Dynamical Mean Field Theory
43
Imaginary part of Al’s DMFT hybridization function for a few components considering zero self-energy. a)Δ00, b)Δ01,
c)Δ10, d)Δ11
• Dynamical mean-field theory (DMFT): commonly used
techniques for solving predicting electronic structure of
correlated systems using impurity solver models.
• DMFT maps a many-body lattice problem to a many-
body local problem with impurity models.
• In DMFT one of the central quantities of interest is the
Green’s function such as
𝐺(𝑘, ꞷ𝑛) = [ꞷ𝑛 + 𝜇 − 𝐻(𝑘) − 𝛴(ꞷ𝑛)]−1
• Spectral function (𝐴) & DMFT hybridization function (𝛥)
𝐴(ꞷ) = −
1
𝜋
𝐼𝑚(𝐺(ꞷ + 𝑖𝛿))
𝑘
𝛥(ꞷ + 𝑖𝛿) = ꞷ − (𝐺)−1
• Next, integrate with quantum impurity solvers
𝛴 = 0