Dear Peter, Laurence, and Xavier:
many thanks for looking into this issue and making suggestions.
The future plan is to go to 128+ atoms supercell for alloys. So the
computational efficiency will be important at that point.
I also tried to eliminate all symmetry operations except for
translation, of course. The structure file is at the bottom. There are
some promising results obtained on a course k-grid.
k-mesh 8x8x8 not shifted, 1 symmetry operation
[oleg@feynman InP-w2k]$ head InP-w2k.epsilon
#
# Lorentzian broadening with gamma= 0.100000 [eV]
# Im(epsilon) shifted by 0.7860 [eV]
# No intraband contributions added
#
# Energy [eV] Re_eps_xx Im_eps_xx Re_eps_yy Im_eps_yy
#
0.013610 0.906000E+01 0.930282E-01 0.906003E+01 0.930290E-01
0.040820 0.906072E+01 0.943965E-01 0.906076E+01 0.943973E-01
0.068030 0.906217E+01 0.958009E-01 0.906220E+01 0.958016E-01
k-mesh 8x8x8 shifted, 1 symmetry operation
[oleg@feynman InP-w2k]$ head InP-w2k.epsilon
#
# Lorentzian broadening with gamma= 0.100000 [eV]
# Im(epsilon) shifted by 0.7860 [eV]
# No intraband contributions added
#
# Energy [eV] Re_eps_xx Im_eps_xx Re_eps_yy Im_eps_yy
#
0.013610 0.869517E+01 0.842315E-01 0.869517E+01 0.842315E-01
0.040820 0.869576E+01 0.853542E-01 0.869576E+01 0.853542E-01
0.068030 0.869695E+01 0.865021E-01 0.869695E+01 0.865021E-01
It seems that both shifted and unshifted mesh could work. I lean toward
an unsifted mesh since the direct gap is at Gamma, so I would prefer to
have it in the k-mesh. Even without symmetry 16x16x16 mesh might be more
computationally efficient than the high-density mesh? The alloy
structure will likely to have no symmetry either.
Going forward, I can try to see how far should the symmetry be reduced.
Next candidate can be a face-centered orthorhombic structure. Any other
thoughts?
Best regards
Oleg
P.S. Here is the structure file
[oleg@feynman InP-w2k]$ cat InP-w2k.struct
InP
F LATTICE,NONEQUIV.ATOMS: 2
MODE OF CALC=RELA unit=ang
11.090240 11.090240 11.090240 90.000000 90.000000 90.000000
ATOM -1: X=0.00000000 Y=0.00000000 Z=0.00000000
MULT= 1 ISPLIT= 8
In NPT= 781 R0=0.00001000 RMT= 2.0000 Z: 49.000
LOCAL ROT MATRIX: 1.0000000 0.0000000 0.0000000
0.0000000 1.0000000 0.0000000
0.0000000 0.0000000 1.0000000
ATOM -2: X=0.25000000 Y=0.25000000 Z=0.25000000
MULT= 1 ISPLIT= 8
P NPT= 781 R0=0.00010000 RMT= 2.0000 Z: 15.000
LOCAL ROT MATRIX: 1.0000000 0.0000000 0.0000000
0.0000000 1.0000000 0.0000000
0.0000000 0.0000000 1.0000000
1 NUMBER OF SYMMETRY OPERATIONS
1 0 0 0.00000000
0 1 0 0.00000000
0 0 1 0.00000000
1
On 2018-09-05 06:10, Peter Blaha wrote:
Dear Oleg,
I looked into the problem and unfortunately I can offer only a partial
solution. I confirm that:
a) The scf cycle gives identical results with or without broken symmetry.
b) The optics gives "wrong" results with broken symmetry.
I inspected the matrix elements and the problem seems to be with
degenerate states (eg. the VBM is 3 fold degenerate at Gamma).
The corresponding momentum matrix elements in cubic case are all the
same (for all 3 eigenvalues and all directions, i.e. Mxx=Myy=Mzz and
M10_13 = M11_13 = M_12_13, where the numbers indicate the band indices
of the transition).
In the non-cubic setup, they are NOT the same, but in my opinion they
are still correct (If you sum over the 3 eigenvalues 10-12, you get the
same result as in the cubic one, but individually I get (for distortions
in z, not in x as you did) M_10_13: (0 0 zz); M_11_13 (zz/2 zz/2 0) and
the same for M_12_13, while in cubic all 3 matrix elements are (zz/3
zz/3 zz/3).
So I concluded the problem is in the tetrahedron method used in joint.
However, I was not able to find a "bug" in SRC_joint. It seems to be
inherent to the method.
The partial fix: Do it "brute force", i.e. increase the number of
k-points until convergence: For instance with k-meshes of
27 27 27 eps1-xx/zz(0.0136eV) 0.117709E+02 0.117134E+02
34 34 34 0.118539E+02 0.118216E+02
200000k (58 58 58): 0.119641E+02 0.119573E+02
200000k cubic setup: 0.119618E+02 0.119618E+02
Obviously, the error in the tetrahedron method due to degenerate
eigenvalues (like at Gamma or other high symmetry points) is reduced the
more "general k-points" are in the mesh and the result converges towards
the cubic result. In addition, eps-1(0) changes anyway from 11.7 to 11.9
with these k-meshes, so are not yet fully converged.
In terms of cpu-time at least for such cells it is not really a problem
to use a 100 100 100 grid (or more).
Best regards
Peter
On 09/03/2018 05:33 AM, Oleg Rubel wrote:
Dear Wien2k community,
I try to compute opto-elastic properties of InP (zinc-blend
structure). It is related to a change of the dielectric constant (real
part) in response to an applied strain. There are no problems with a
response to a hydrostatic strain, and results agree well with
experiments. A problem occurs with a uniaxial strain (strained along
X-axis only by 0.05%). Computed change in the dielectric constant is
too large (~ an order of magnitude).
Trying to trace back the problem, I did the following:
First, I initialize a tetragonaly-distorted zinc-blend structure
(init_lapw -b -vxc 19 -ecut -6.5 -numk 800) with the following lattice
parameters
F LATTICE,NONEQUIV.ATOMS: 2
MODE OF CALC=RELA unit=ang
11.095785 11.090240 11.090240 90.000000 90.000000 90.000000
Then I set the lattice parameters back to the cubic lattice
F LATTICE,NONEQUIV.ATOMS: 2
MODE OF CALC=RELA unit=ang
11.090240 11.090240 11.090240 90.000000 90.000000 90.000000
and rerun (x dstart). This allows me to preserve the symmetry of a
distorted structure (see the structure file below).
Next, I run SCF (run_lapw -ec 0.00001 -cc 0.0001) and optics with
20x20x20 k-mesh. The results for Re and Im parts of the dielectric
constant are here:
[oleg@feynman InP-w2k]$ head InP-w2k.epsilon
#
# Lorentzian broadening with gamma= 0.100000 [eV]
# Im(epsilon) shifted by 0.7860 [eV]
# No intraband contributions added
#
# Energy [eV] Re_eps_xx Im_eps_xx Re_eps_yy Im_eps_yy
#
0.013610 0.940850E+01 0.988634E-01 0.947674E+01 0.100908E+00
0.040820 0.940928E+01 0.100340E+00 0.947756E+01 0.102453E+00
0.068030 0.941084E+01 0.101855E+00 0.947919E+01 0.104042E+00
It seems that the symmetry is broken, which causes later problems with
opto-elastic coefficients as change of 0.07 in the second decimal
point of Re_eps for such a small strain is too much.
Once again, there are no problems when the strain tensor does not
break the zinc-blend cubic symmetry.
Any thoughts are highly appreciated.
Thank you in advance
Oleg
_______________________________________________
Wien mailing list
Wien@zeus.theochem.tuwien.ac.at
http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
SEARCH the MAILING-LIST at:
http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html
