Stress and variable cell optimization in...
Transcript of Stress and variable cell optimization in...
![Page 1: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/1.jpg)
Stress and variable cell optimization
in OpenMX
• Purpose of the study
• Stress tensor in OpenMX
• Approximate Hessian by Schlegel
• Benchmark calculations
• Optimization of enthalpy
• Summary
Taisuke Ozaki (ISSP, Univ. of Tokyo)
Yoshinori Shiihara (Toyota Tech. Inst.)
Masanobu Miyata (JAIST)
Nov. 25th, The 2nd OpenMX developer’s meeting in KAIST
![Page 2: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/2.jpg)
Purpose of the study
• Full optimization of systems including internal
coordinates and cell vectors
• Acceleration of optimization: reduction of iterative
steps
• Optimization of the enthalpy: variable cell
optimization under pressure
• Molecular dynamics under NPT ensemble
Done
Planned
![Page 3: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/3.jpg)
Stress tensor and derivatives w.r.t cell vectors
a1=(a11,a12,a13)
a2=(a21,a22,a23)
a3=(a31,a32,a33)
a’1=(a’11,a’12,a’13)
a’2=(a’21,a’22,a’23)
a’3=(a’31,a’32,a’33)
' (I ε) r r
i
ij ij j
E E Eb
a a
Strain tensor ε scales the
Cartesian coordinate as
Then, the stress tensor
ij
E
a
E
can be related the energy derivative
w.r.t. cell vectors by
1b a
where
![Page 4: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/4.jpg)
Stress tensor in OpenMX
(NL)
na ec δee XC scctot kinE E E E E E E
Thus, at least there are six contributions to stress tensor.
• The terms are decomposed to derivatives of matrix elements and overlap stress,
leading to rather straightforward analytic calculations.
• The term is analytically evaluated in reciprocal space.
• The term is analytically evaluated in real space with a carefully derived formula.
The computational time is almost the same as that for the force calculation.
In OpenMX, the total energy is defined by
![Page 5: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/5.jpg)
Stress tensor for Ekin, Ena, and Eec
( )
,kin ˆ( ) ( )n
i j
i i j i n
n i j
ET
R
r t r t R
( )
,ˆ( ) ( )n
i j i i j i n
n i j
T
Rr t r t R
,ˆ ˆ( ) ( ) ( ) ( )i i j i n i i j i n ij n
i
T T tt
r t r t R r t r t R
The derivative of Ekin is given by
The latter derivatives can be transformed to the derivatives w.r.t. Cartesian coordinates:
( )
( )
, ,
n
ni j
i j ij n
n i j i
SE t
t
R
R
The former derivatives can be transformed to the overlap stress tensor:( )
, ˆ( ) ( )n
i j
i i j i n
n i j
T
R
r t r t R
,ij n i j n t t t R
where
The energy terms, Ena and Eec, can also be evaluated in a similar way.
![Page 6: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/6.jpg)
Stress tensor for Eδee
ee HH H
( )1 ( ) 1( ) ( ) ( ) ( )
2 2
E Vnn V d V d n d
rr
r r r r r r r
The derivative of Eδee is given by
H
1 ( )( )
2
nV d
r
r r
The second term is given by
H ( )1( )
2
Vn d
r
r r
The third term is given by
![Page 7: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/7.jpg)
Stress tensor for Exc
( ) ( )
xc ,PCC termp p
p
p p
n nV v V A
x
The derivative of Exc is given by
The second term contributes the overlap stress tensor, and third term can
be evaluated as( ) ( )
xc,( ) ( )
| |
| |
p p p p
p p
p p pp p
n n n nfV V A V A
n n x
( )
xc
( ) ( )
| |
| |
p
p
p p
nfA
n n
where
( ) ( ) ( )
XC xc xcxc ( ) ( ) ( )
| |PCC term
| |
p p p
p pp p p
n n nE f fE V V
n n n
The last term is given by
![Page 8: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/8.jpg)
Approximate Hessian by Schlegel
2
1(| |)
2i n j
i Rn j
V f r R r
3( )
AF
r B
Schlegel proposed a way of constructing an
approximate Hessian. A force constant for every pair
of elements is fitted to the following formula, where
dataset were constructed by B3LYP calculations.H.B. Schlegel, Theoret. Chim. Acta (Berl.) 66, 333
(1984); J.M. Wittbrodt and H.B. Schlegel, J. Mol. Struc.
(Theochem) 398-399, 55 (1997).
Suppose the total energy is given by the sum of pairwise potentials. Then, the derivatives
lead to the following relation:
H BFwhere B is the B-matrix of Wilson, H is the approximate Hessian in Cartesian coordinate.
![Page 9: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/9.jpg)
Benchmark of the approximate Hessian
in OpenMX
For both molecules and bulks, it is found that the Schlegel’s
method improves the convergence substantially.
Molecules Bulks
![Page 10: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/10.jpg)
Variable cell optimization
Initial Hessian: Schlegel’s method
Preconditioning: RMM-DIIS
Hessian update: BFGS
Update of positions: Rational function (RF)
RF method
It is very important to construct the initial Hessian including
internal coordinates, cell vectors, and the cross term for fast and
stable convergence.
Int Int CellH BF
Cell Int Cell
![Page 11: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/11.jpg)
Benchmark calculations of RFC5
For 785 crystals (mostly sulfides) , the full optimization by RFC5 were
performed by Mr. Miyata, Ph.D student in JAIST, as computational
screening in searching good thermoelectric materials
The optimization criterion: 10-4 Hartree/bohr
The histogram shows the number of systems among
785 systems as a function of the number of iterations
to achieve the convergence
![Page 12: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/12.jpg)
Optimization of the enthalpy
H E pV
H E V Ep pV
Under an external pressure p, the
structural optimization can be performed
by minimizing the enthalpy defined with
The stress tensor is easily calculated by
La3Si6N11: Ce2c
History of optimization
10 GPa
![Page 13: Stress and variable cell optimization in OpenMXt-ozaki.issp.u-tokyo.ac.jp/meeting16/OMX-Ozaki-2016Nov-7.pdfStress and variable cell optimization in OpenMX •Purpose of the study •Stress](https://reader031.fdocuments.us/reader031/viewer/2022020108/5b1c55667f8b9a41258fd168/html5/thumbnails/13.jpg)
Summary
• We have derived an analytic formula of stress tensor and
implemented the stress tensor in OpenMX.
• Acceleration of optimization has been achieved by introducing
an approximate Hessian by Schlegel, which is effective to reduce
the number of iterative steps.
• It is found from benchmark calculations of 785 systems that most
of systems converge around 20-40 iterations.
• Optimization of the enthalpy was implemented, enabling variable
cell optimization under pressure.