Upload
sameh-lotfy
View
223
Download
0
Embed Size (px)
Citation preview
8/12/2019 (14) Iterative Minimization of Energy Part 2
1/20
POSTECH
Special Lecture on Density Functional Theory:
(14) Iterative Minimization of Energy by theConjugate-Gradient (CG) Method
by Prof. Hyun M. Jang Dept. of Materials Science and Engineering, and Division of Advanced
Materials Science, Pohang University of Science and Technology(POSTECH), Republic of Korea.
also at Dept. of Physics, Pohang University of Science and Technology(POSTECH), Republic of Korea.
8/12/2019 (14) Iterative Minimization of Energy Part 2
2/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
Ab initio total-energy pseudo-potential calculations by conventionalmatrix diagonalization technique require significant amounts ofcomputer time, even for systems containing a few atoms in the unit
cell. The Hamiltonian matrix for each of the kpoints in the IBZ mustbe constructed, as discussed in the Kohn-Sham (K-S) equation in aplane-wave basis, and diagonalized to obtain the K-S eigenstates.
This procedure should be repeated until the charge density for of
the (n-1)th iteration is self-consistent with that of the nth iteration.
The cost of matrix diagonalization increases as the 3rd power of thenumber of plane-wave basis states. As a result, conventional matrix
diagonalization techniques are restricted to the order of 10 atoms inthe unit cell (order of 1,000 plane-wave basis states).
In this chapter, an efficient method is introduced that allows direct
minimization of the Kohn-Sham energy functional based on the
conjugate-gradient (CG) method.
H
8/12/2019 (14) Iterative Minimization of Energy Part 2
3/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
* Introduction to Conjugate-Gradient (CG) Method
Two general methods are commonly used to locate the minimum of
function F(x), where x is a vector in the mulltidimensional space: (i) the
method of steepest descent and (ii) the conjugate-gradient (CG) method.
In the absence of any information about
the function F(x), the optimum direction to
move from the point x1 to minimize the
function is just the steepest-descent
direction g1 given by
where is the gradient operator acting on
the vector x1
to give the steepest-descentdirection g1. Although each iteration of the
steepest-descent algorithm moves the trial
vector towards the minimum of the
function, there is no guarantee that the
minimum will be reached in a finite numberof iterations (See the figure at rhs).
G
)1(.......... 1xx
1x
xg
1G
=
=
8/12/2019 (14) Iterative Minimization of Energy Part 2
4/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
In many cases a very large number of steepest-descent iterations isneeded to get close to the minimum.
Although it may seem surprising, there is a faster way to reach theminimum than to follow the downhill steepest-descent direction.
Let d1be the minimization direction ofF(x) at x1 and d2be the
subsequent minimization direction at x2, where x2 = x1+b1d1. Then, it
can be shown that the following relationship should be hold if d
1
andd2 are to be independent (because each minimization step is indepen-
dent of the previous steps):
This is the minimization condition that the directions d1 and d2be
conjugate to each other and can be generalized to
The conjugate-gradient (CG) technique provides a simple and
effective procedure for implementation of such a minimization app-
)2(...................0 12 == dddd 21 GG
)3(..............................0 mnforG = mn dd
8/12/2019 (14) Iterative Minimization of Energy Part 2
5/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
roach. The initial direction is taken to be the negative of the gradient(steepest-descent direction) at the starting point. A subsequentconjugate direction is then constructed from a linear combination
of the new gradient and the previous direction that minimizedF(x). In a two-dimensionalproblem, it is clear that one would needonly two conjugate directions, and this would be sufficient to span the
space and arrive at the minimum in just two steps, as described in the
figure of Page 2.
The above descriptions can be generalized by the followingalgorithm:
In Eq. (4), dm is the (search) conjugate direction in the mth iterationand gm is the steepest-descent vector or direction in the mth iteration.
(Ref.) Appendix L of R. M. Martins book.
)5(................0
)4(.............................................
1
=
+=
withwhere m
m
1m1m
mm
1mmm
gggg
dgd
1221222 ggdgd +=+=
8/12/2019 (14) Iterative Minimization of Energy Part 2
6/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
* Orthogonality Constraints in the Conjugate-Gradient Method
In this section, a computational technique that can overcome the
problem associated with large supercell sizes and large plane-wavekinetic-energy cutoffs is introduced. This technique adopts the
conjugate-gradient (CG) approach, with the proper preconditioning, tominimize directly the Kohn-Sham energy functional.
In the case of total energy calculations, the Kohn-Sham energyfunctionalEtakes the place of the functionF, the wavefunctionstake the place of the vector x, and the Kohn-Sham Hamiltonian isthe relevant gradient operator
The steepest-descent direction which satisfies the orthogonalityconstraints of is given by
where the superscript m labels the iteration number and the super-
i
.G
)mi
{ }i
)6(..............................j
m
i
ij
j
m
i
m
i =
8/12/2019 (14) Iterative Minimization of Energy Part 2
7/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
script i is the band index. There is no iteration index on the wave-
functions in Eq. (6) because do not vary during iterations for bandi. is the steepest-descent direction obtained without considering
the orthogonality constraints with the definition
Eq. (6) can be obtained by requiring the following orthogonality:
If the search direction for band iwere not orthogonal to the wave-functions of all the other bands, all
of the wave-functions would have to change during each iteration inorder to maintain the constraints of orthogonality.
(Ref.) (i) M. P. Teter, M. C. Payne, and D. C. Allan, Solution of Schrdingerequation for large systems, Phys. Rev. B 40, 12255-12263 (1989). (ii) M. C. Payne,M. P. Teter, D. C. Allan, T. A. Arias, and J. D. Joannopoulos, Iterative minimization
technique for ab initio total-energy calculations: molecular dynamics and conjugategradients, Reviews of Modern Physics 64, 1045-1097 (1992).
m
i
j
)7(..... mi
m
i
m
i
m
i
m
i
m
i
m
i HHHH ==
)8(.....0===
m
il
m
iljlj
m
ij
m
il
m
il .0,0 == m
ilji thenif
.im
i
m
i =
8/12/2019 (14) Iterative Minimization of Energy Part 2
8/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
* Preconditioning Vector
The relation between the error in the wavefunction, , and thesteepestdescent direction is obtained by expanding in terms
of the eigen-states of the Kohn-Sham Hamiltonian, namely,
The steepest-descent vector is obtained by substituting Eq. (9) into
Eq. (7).
where is the eigenvalue associated with the eigen-state
The Kohn-Sham Hamiltonian has a broad spectrum of eigenvalues(i.e., a broad spectrum of the steepest descent vectors). This leads to
poor convergence in a conjugate-gradient calculation. The techniqueof preconditioning is proposed to improve the rate of convergence.
The preconditioning technique involves multiplying the steepest-descent vector by a preconditioning matrixK to produce a precondi-
i
i
i
)9(...........................................==
iii c
( ) )10(.................
== iiiii ccH
(e.g., plane wave)
8/12/2019 (14) Iterative Minimization of Energy Part 2
9/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
tioned steepest-descent vector that more accurately represents the
error vector, as illustrated in the figure presented below. Thus, onecan obtain an essentially eigenvalue-independent steepest-descent
vector by exploiting this technique.
The preconditioned steepest-descent
vectors that accurately represent
can be obtained by multiplying thesteepest-descent vectors by an ortho-
gonalized preconditioning matrix,
namely,
where the matrix element of the
preconditioning matrixKis
expressed by the following equation:
)11(....................mimi K =
i
( ) ., iii as
8/12/2019 (14) Iterative Minimization of Energy Part 2
10/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
in Eq. (12) is defined by
where is the kinetic energy of the stateThe matrix elements asymptotically approach 1/{2(+1)}.
This factor (i.e., ) thus causes all of the large wave-vector
(i.e., large ) to converge at nearly the same rate, as shown in thefigure presented in the previous page.
The preconditioned steepest-descent vector as presented in Eq. (11)
is not orthogonal to all the bands. On the analogy of Eq. (6), thepreconditioned steepest-allowed-descent vector that is orthogonalto all the bands is calculated as
where a prime signifies a descent vector orthogonal to all the bands.
( ))12(.............
168121827
8121827432
32
++++
+++= GG,GG,K
)13(................1
2
22
m
iTm
Gk+h
m
i
m
i
m
im
T
= 2
2
2
h .mi
GG,
1GG,
)14(................................j
j
m
ij
m
i
m
i =
8/12/2019 (14) Iterative Minimization of Energy Part 2
11/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
* Conjugate Directions
The conjugate-gradient direction can be constructed out of steepest-descent vectors according to
where is the preconditioned steepest-descent direction. Eq. (15) isbased on Eq. (4), namely, where gm is the steepest-
descent vector in the mth iteration. is given by
where The conjugate direction generated by Eq. (15) will notbe orthogonal to wave-function of the present band. A furtherorthogonalization to the present band can be done by
)15(...........,. 12221iiii
m
i
m
i
m
i
m
i ge +=+=
m
i
,1mmm dgd += mmi
)16(...............................1111
=
=
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
im
i
.0
1 =i m
i
)17(.................................mi
m
i
m
i
m
i
m
i =
0== m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
8/12/2019 (14) Iterative Minimization of Energy Part 2
12/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
Then, a normalized conjugate direction can be computed by
is then orthogonal to all the bands, because of Eq. (17).
* Search for the Energy Minimum
Define the following linear combination as a trial eigenvector(wave-function) for the (m+1)th iteration:
According to Eq. (19), is a measure of the deviation from thenormalized conjugate direction, Thus, Eq. (19) is orthogonal to
all the other bands satisfying the orthogonality constraints.Teteret al. [Phys. Rev. B 40, 12255-12263 (1989)]proposed the following
-dependent Kohn-Sham energy expression to locate the minimum of
m
i
m
i
)18(...................................2/1m
i
m
i
m
im
i
=
)19(............)(sincos1
realism
i
m
i
m
i +=+
2/ .mi
),( i
j
jThe minimum will occur at xm+1 along
the conjugate direction dm according
to xm+1 = xm + bmdm.
8/12/2019 (14) Iterative Minimization of Energy Part 2
13/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
the K-S energy functional:
WritingEKSas a function of
is equivalent to writingEKS as afunction of Three pieces of information are required to evaluate
the three unknowns in Eq. (20), namely,Eavg,A1, andB1.
The conjugate-gradient technique requires that the value of that
minimizes the Kohn-Sham energy functional be found. To do this, letus considerE() in details.
where
.m
i
)20(.......)2sin()2cos()(11
avg
++=
)21(......................)()(
)()()()(21
)()()(
1
1
1
333
1
1
*1
rrr,
rrrrrr
xcH
m
i
m
i
xc
m
i
m
i
EEH
rdnrdrdnn
dHE
++=
+
+=
++
++
).22(............................1 ext
VTH +=
The 2nd and 3rd terms of
Eq. (20) reflect the
variation of E during the
iteration of i.
8/12/2019 (14) Iterative Minimization of Energy Part 2
14/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
According to Eq. (21), we also have to evaluate term.
{ } { }
{ } { }
{ } { }
{ } )24(...cossinsincos
sincoscossin)(
,
)23(............sincossincos
)(
111*
1*
1
**
1
**
0
1
1
**
)19.(
1
1
1
1
mi
mi
mi
mi
mi
mi
mi
mi
o
m
i
m
io
m
i
m
i
o
m
i
m
io
m
i
m
i
m
i
m
i
m
i
m
i
Eq
m
i
m
i
HHdHH
dH
dHE
Then
dH
HELet
+= +=
+ ++
+ +=
+ +
=
=
++
0=
+
xcH
EE
{ }
)26(.....sincossincos)(
)25(.....)()()()( 000
m
i
m
i
m
i
m
i
xcH
xcHxcH
nand
nVVnnE
nEEE
++=
+=
+
=
+
=
r
rrrr
8/12/2019 (14) Iterative Minimization of Energy Part 2
15/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
has been computed to determine the steepest-descent vector
[Eq. (7)]. Thus, the value of Eq. (29) can be computed readily.B1 can be determined.
{ } )29(...............................2Re2
)28(..........
1)20.(
11
000
1
0
0
BH
VVHVVH
EEEE
VVVVEE
Eq
m
i
m
i
m
ixcH
m
i
m
ixcH
m
i
xcH
m
ixcH
m
i
m
ixcH
m
i
xcH
==
+++++=
+
+
=
+++=
+
====
=
{ }{ }
{ }{ }
{ } )27(...............0cos0sin0sin0cos
0sin0cos0cos0sin)(
**
**
**
0
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
d
d
dn
+=+=
+ ++
+ +=
=
r
m
iH
8/12/2019 (14) Iterative Minimization of Energy Part 2
16/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
To determine the two other parameters (Eavg,A1), Teteret al.proposed analytic method that utilizes the K-S energy at a secondvalue of The sampling point of the second point should
be far enough from to avoid rounding errors but not so far fromthe origin. It has been found that computing the K-S energy at thepoint gives reliable results.
From Eq. (20), one can obtain the following expression ofEavg
for = /300.
As shown in Eq. (29), B1 is given by
( ).0
0=
300/=
)31(........)300/2cos(1
)300/2sin(21)300/()0(
4
1
)30(.....)300/2cos(1
)300/2cos()0()300/2sin(2
1)300/(
0
0
2
2
1
0
+
=
=
=
=
=
EEEE
Aand
EEE
avg
.2
101 =
= B
8/12/2019 (14) Iterative Minimization of Energy Part 2
17/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
Once the three parameters (Eavg,A1, andB1) have been determined,
the value of that minimizes the Kohn-Sham energy functional atthe mth iteration can be evaluated. Then, the following relation should
be hold for the minimum K-S energy.
The value of that lies in the range is the required
value.
According to Eq. (32), one has to evaluate the second derivative of
Eto correctly locate
)32(.....
2
1tan
2
1tan
2
12tan0
02
2
01
1
11
min
1
1
min)20.(
=
===
=
=
E
E
A
B
A
BEEq
min 2/0
8/12/2019 (14) Iterative Minimization of Energy Part 2
18/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
The required value of is then determined using Eq. (32). Thewave-function used to start the next iteration of the conjugate-gradient
procedure, , is given by [See Eq. (19)].
The new generates a different charge density from the density
generated by previous , and so the electronic potentials in the K-SHamiltonian must be updated before commencing the next iteration.
min
{ }
{ }
{ }
{ }[ ] )33(......................)(
)()(Re2
)()(Re2*1
2
2
32*
2
2
0
3*
2
2
02
2
rdn
Vf
rdeG
ef
HH
termsXCHartreeHHE
xcm
i
m
i
m
i
m
i
i
o
m
i
m
i
m
i
m
i
m
i
m
i
m
i
m
i
rrr
rrG
rG
+
+
=
=++=
=
)34(................................sincosminmin
1 mi
m
i
m
i +=+
1+m
i
1+m
i
m
i
8/12/2019 (14) Iterative Minimization of Energy Part 2
19/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method
Trial Wavefunction for a particular band,m
i
Calculate the steepest-descent vector,
* Computational Procedure of the CG Method
m
i
m
i
m
i H =
Orthogonalize to all bands, jm
iij
j
m
i
m
i =
Compute the preconditioning steepest-descent vector, mi
m
i K =
Orthogonalize to all bands, jij
m
ij
m
i
m
i
m
i
m
i
m
i =
Determine the conjugate direction,
Orthogonalize to the present band and normalize,
Calculate the K-S energy at initial value of ,
Calculate the value of that minimizes the K-S energy functional.
Construct a new trial wavefunction,
1+= mi
m
i
m
i
m
i
m
i
)2sin()2cos()(11
avg
++=
minmin
1 sincos mi
m
i
m
i +=+
Repeat
until
converged.
8/12/2019 (14) Iterative Minimization of Energy Part 2
20/20
Iterative Minimization of the Kohn-Sham Energyby the Conjugate-Gradient Method