Accelerated Quantum Molecular Dynamics · occ. | i|2! ⇢ sc SCF SCF SCF SCF U(R; ) U(R; sc)...
Transcript of Accelerated Quantum Molecular Dynamics · occ. | i|2! ⇢ sc SCF SCF SCF SCF U(R; ) U(R; sc)...
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Accelerated Quantum Molecular Dynamics
Enrique Martinez, Christian Negre, Marc J. Cawkwell, Danny Perez, Arthur F. Voter and Anders M. N. Niklasson
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Outline
§ Quantum MD
• Current approaches
• Challenges
§ Extended Lagrangian Born-Oppenheimer MD
§ Accelerated MD
§ Coupling XL-BOMD and AMD
• Vacancies in graphene
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Integrate the equations of motion of classical molecular trajectories
with the forces calculated on the fly from a self-consistent quantum mechanical description of the electronic structure:
MIRI = �⇥U(R; �sc)
⇥RI
H[�]�i = ⇥i�i
⇢ =X
occ.
|�i|2! ⇢sc
SCF SCFSCF
SCF
U(R; �)
U(R; �sc)
Born-Oppenheimer MD
Goal: To develop a unique computationalcapability based on a new generationquantum based molecular dynamics thatovercomes current limitations allowingdesign and prediction of materials andprocesses on time and length scalesmultiple orders of magnitude beyondcurrent capacity.#SCF�O(N3)
Quantum based MD
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
New Generation Quantum Mechanical Molecular Dynamics
XL-BOMD
Car-Parrinello MD (1985)Born-Oppenheimer MD (1973)
+ Accurate and reliable+ General, any material + Long time steps- Energy drift- Expensive, SCF cost
- Parameter dependence- Material dependent- Short time steps+ Small energy drift+ Fast, no SCF required
The best of both worlds
Quantum based MD
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
System Size (N)
Wal
l-Clo
ck T
ime
/ MD
Tim
e St
epO(N3) DiagonalizationO(N) Regular linear scalingO(N) Low pre-factor
Significant pre-factor reduction is required for practical
large scale QMD!
Quantum Molecular Dynamics
The Challenge
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D Slide 6
Hi�,j⇥ [q] = hi�,j⇥ +X
l
ql�il⇥ij⇥�⇥
“Self-consistent-charge density-functional tight-binding method for simulations of complex materials properties”
M. Elstner et al. Phys. Rev. B 58, 7269 (1998) +1,400 citations
UDFTB(R; ⇥) = 2Tr[⇥h] +1
2
X
i,j
qiqj�ij + Epair[R]
qi = 2X
�
�i�,i� ⇥ = � (µI �H[q])
Maybe the simplest and most efficient approach to electronic structure theory that still covers the relevant
features of “exact” ab initio Kohn-Sham DFT.
Typically two orders of magnitude in speed up (w.r.t. DFT)!
LATTE: Self-Consistent Charge Density Functional based Tight-Binding Theory
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
SCFSCF
SCF
Potential energy surface at the SCF optimized electronic ground state
Born-Oppenheimer MD
Potential energy surface of a non-optimized electronic state
Car-Parrinello MDDtBO
dtCP << DtBO
Potential energy surface of a shadow H is very closeto the exact SCF optimized electronic ground state
0New Fast “SCF-free” GXL-BOMD
DtXL-BO
Niklasson & Cawkwell, JCP 141, 164123 (2014)
SCF-free Generalized XL-BOMD
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
SCF-free Generalized XL-BOMD
Slide 8
Niklasson & Cawkwell, JCP 141, 164123 (2014)
MIRI = � �U(R, n)
�RI
����n
n(r) = �⇥2
ZK(r, r0) (�[n](r0)� n(r0)) dr0
U(R, n) = min�
(E(1)
DFT
(R, ⇥, n)
����� ⇥ =X
occ
|�i|2, ��i|�j⇥ = �ij
)+ V
nn
(R)
E(1)DFT(R, ⇥, n) = EDFT(R, n) +
Z�EDFT(R, n)
�⇥
�����=n
(⇥[n](r)� n(r)) dr
K(r, r0) ⇠✓�⇥[n](r
�n(r0)� �(r� r0
◆�1
⇥[n] = argmin�
(E(1)
DFT
(R, ⇥, n)
����� ⇥ =X
occ
|�i|2, ��i|�j⇥ = �ij
)
ETot
=1
2
X
I
MIR2
I + U(R, n)
LShadowXBO =
X
I
MIR2I
2� U(R, n) +
µ
2
Zn2dr� µ⇥2
2
Z(�� n)†K†K(�� n)dr
LShadowXBO = LKS
BO, {� BO approx. lim� ⇥ ⇤, limµ ⇥ 0, limµ� ⇥ constant}
Shadow Hamiltonian
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
SCF-free Generalized XL-BOMD Canonical Ensemble
Slide 9
200
250
300
350
400
T (
K)
a)
−622.4
−622.3
−622.2
−622.1
−622.0
0 2 4 6 8 10
ES +
EN
(e
V)
time (ps)
b) BO 6 diag/step
BO 9 diag/step
XLBO 1 diag/step
200
250
300
350
400
T (
K)
a)
−622.4
−622.0
−621.6
−621.2
−620.8
0 2 4 6 8 10
Energ
y (e
V)
time (ps)
b)
O(N3)
BO 6 diag/step
BO 9 diag/step
XLBO 1 diag/step
Isocyanic acid (HNCO)24
NVE ensembleNVT ensemble Nose thermostat
Is the XL-BOMD compatible with the thermostats? Does it help?
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D Slide 10
SCF-free Generalized XL-BOMD Canonical Ensemble
350
400
450
500
550
600
650
700
1 2 3 4 5 6 ∞T (
K) Born−OppenheimerXLBO
Un
sta
ble
a)(HNCO)24
Ttarget = 500 K
350
400
450
500
550
600
650
1 2 3 4 5 6 ∞ T
Diagonalizations per step
Born−OppenheimerXLBO
Un
sta
ble
b) Andersen Langevin
Nose(CH3NO2)32
260
280
300
320
340
T (
K)
Unstable
a) AndersenLangevin
Nose
(HNCO)24 Ttarget = 300 K
400
450
500
550
600
0 0.2 0.4 0.6 0.8 1.0 1.2
time step (fs)
b) nitromethane Ttarget = 500 K
Average temperature and the square-root of its second moment
XL-BOMD converges faster with the right fluctuations
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
XL-BOMD + Accelerated Molecular Dynamics
Slide 11
a
bg
Infrequent Events
Can we accelerate the transition without perturbing the
probabilities for the system to escape? AMD
Accelerated Molecular DynamicsHyperdynamics
Parallel Replica Dynamics
Temperature Accelerated Dynamics
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Parallel Replica Dynamics
Slide 12
The properties of the exponential allow us to parallelize time, by having many processors seek the first escape event. The procedure:
Must detect every transition.
dephase allow correlated events
parallel time
AFV, Phys. Rev. B 57, 13985 (1998).
add all these times together (tsum) when first event occurs on any processor
(overhead)
tcorr
tcorr
replicate
Good parallel efficiency if trxn / nproc >> 2tcorr
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D Slide 13
Master
Replica 1 Replica 2 Replica n…
Python script
LATTE: Tight Binding Order NParallel Implementation MultiThreads
AMDF(BES program)
XL-BOMD + Accelerated Molecular Dynamics
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Application: Vacancies in graphene
Slide 14
We have studied 1 and 2 vacancy diffusion
Temperature distribution of a graphene sheet containing 1V at 2500 and 3000 K in the NVT using Langevin Dynamics
run for 100 ps
a) b)
2500 3000 3500 4000 4500
Ttarget = 3000 K
2996 K3058 K
3281 K
BO−1SCFunstable
T (K)
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
2000 2500 3000 3500 4000 4500
Ttarget = 2500 K2493 K2582 K
2681 K
3631 K
frequ
ency
T (K)
BO−1SCFBO−3SCF
BO−5SCFXLBO−1SCF
We have analyzed the normal modes to pick the dephasing time ~5ps. There are some slow modes perpendicular to the sheet plane
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Application: Vacancies in graphene
Slide 15
1V diffusivities in NVT ensemble ParRep-LATTE, reaching on the order of hundreds of ns
10−10
10−9
10−8
10−7
10−6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
D = ν λ2 exp(−∆E/kBT)λ = 1.54 10−10 m
ν = 9.44 1012 s−1
∆E = 0.54 eV
ν = 1.17 1015 s−1
∆E = 1.16 eV
DV
(m2 /s
)1000/T (K−1)
Migration barriers in agreement with recent DFT calculations
Competing events at higher temperatures
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D Slide 16
2V configurations at 3500 K in NVT ensemble ParRep-LATTE
Net translation of the center of mass
Application: Vacancies in graphene
4−55−7−9
5−77−5
5−8−5
555−777
5555−6−7777
0.0 100 5.0 10−9 1.0 10−8 1.5 10−8 2.0 10−8 2.5 10−8 3.0 10−8
t (s)
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D Slide 17
Application: Vacancies in graphene
5-8-5 555-777 5555-6-7777 2x(57) 4-55-7-9
Experiments 50.3% 14.1% 18.8% 1.8% 0%
Simulation 20.7% 65.2% 12.7% 0.6% 0.02%
Mobility and configurational distribution compatible with
experimentsto
tal m
igra
tion
leng
th d
istri
butio
n
x m
igra
tion
leng
th d
istri
butio
n
y m
igra
tion
leng
th d
istri
butio
n
a) b)
c) d)
5-8-5 5-8-5translations
5-8-5 555-777 555-777 5555-6-7777 transitions
Bond strechingRotations
−2
0
2
4
6
8
6 7 8 9 10 11 12 13 14
T = 3500 Kt = 27.87 ns
initial
final
y (Å
)
x (Å)
00.050.1
0.150.2
0.250.3
0.350.4
0 0.5 1 1.5 2 2.5 3d (Å)
0
0.05
0.1
0.15
0.2
0.25
−3 −2 −1 0 1 2 3�x (Å)
0
0.05
0.1
0.15
0.2
0.25
−3 −2 −1 0 1 2 3�y (Å)
Operated by Los Alamos National Security, LLC for the U.S. Department of Energy’s NNSA
U N C L A S S I F I E D
Conclusions
n We have developed a tool coupling XL-BOMD and AMD that reaches ~2-3 orders of magnitude speedup with respect to traditional methods
n We have calculated single vacancy mobilities in graphene reaching hundreds of ns
n We have also studied the very stable 2V-complex migration
Slide 18