Fine q-grid convergence

[Nandan]

This problem seems to continue for me.

I updated the input so that it looks as follows.
On a separte note, I ran all the Examples in the EPW directory and they completed without error.
I could use help with figuring out this issue.


--
&inputepw
prefix = 'gaas'
amass(1) = 69.72300
amass(2) = 74.92160
outdir = './'

! ephwrite = .true.

epbwrite = .false.
epbread = .false.
epwwrite = .false.
epwread = .true.

elph = .true.
kmaps = .true.
etf_mem = 1

nbndsub = 12
nbndskip = 0

dis_win_max = 29
dis_froz_max= 14
dis_froz_min= -8

restart = .true.
restart_freq= 10000

wannierize = .false.
lpolar = .true.
num_iter = 5000
iprint = 2
proj(1) = 'random'
proj(2) = 'As:sp3'

! efermi_read = .false.
! fermi_energy= 5.4304

elecselfen = .true.
phonselfen = .false.

band_plot = .true.
parallel_k = .true.
parallel_q = .false.

fsthick = 25
eptemp = 300.d0
degaussw = 0.010

dvscf_dir = '../phonons/save'
filukk = './gaas.ukk'
filkf = './LGX500kpts.pwscf'
! filqf = '/u/ntandon/source/kpaths_FCC/LGX-100q.pwscf'
! filkf = '/u/ntandon/source/kpaths_FCC/LGX-25K.pwscf'

nk1 = 6
nk2 = 6
nk3 = 6

nq1 = 6
nq2 = 6
nq3 = 6

! nkf1 = 20
! nkf2 = 20
! nkf3 = 20
!
! nqf1 = 50
! nqf2 = 50
! nqf3 = 50

! rand_k = .true.
! rand_nk = 8000

!mp_mesh_k = .true.
rand_q = .true.
rand_nq = 1000000

/
16 cartesian
0.0000000 0.0000000 0.0000000 0.0092593
-0.1666667 0.1666667 -0.1666667 0.0740741
-0.3333333 0.3333333 -0.3333333 0.0740741
0.5000000 -0.5000000 0.5000000 0.0370370
0.0000000 0.3333333 0.0000000 0.0555556
-0.1666667 0.5000000 -0.1666667 0.2222222
0.6666667 -0.3333333 0.6666667 0.2222222
0.5000000 -0.1666667 0.5000000 0.2222222
0.3333333 0.0000000 0.3333333 0.1111111
0.0000000 0.6666667 0.0000000 0.0555556
0.8333333 -0.1666667 0.8333333 0.2222222
0.6666667 -0.0000000 0.6666667 0.1111111
0.0000000 -1.0000000 0.0000000 0.0277778
0.6666667 -0.3333333 1.0000000 0.2222222
0.5000000 -0.1666667 0.8333333 0.2222222
-0.3333333 -1.0000000 0.0000000 0.1111111


Thanks and regards,

Nandan.

[sponce]

Hello,

Except from the fact that I would use something like only
proj(1) = 'Ga:sp3'

I never use "random" projection as it is typically quite bad.

How fast are your wannier function converging (from gaas.wout file) ?

Are all the decay.X files correctly decaying ?

Apart from this, I do not see anything else that seems wrong.

Best wishes,
Samuel

[Nandan]

Hi Samuel,

I finally finished some more tests but my results are still not as expected.

I tested two cases with different number of Wannier functions: https://drive.google.com/file/d/1CU89Zz ... sp=sharing
The decay of H and dynmat in both cases is almost identical. The decay of the electronic part is very different but better for the basis of 8 Wannier functions.

In both cases I still observe my original problem of Im(Sigma) values increases by order of magnitude for grids from 1x10^5 to 1x10^6.

==================
8 Wannier functions
==================
1.begin projections
Ga:sp3
As:sp3
end projections
num_wann 8
iprint 2
dis_win_min -1000.000
dis_win_max 20.000
dis_froz_min -1000.000
dis_froz_max 12.000
num_iter 5000
----------------------------------------

==================
12 Wannier functions
==================
2. begin projections
random
Ga:sp3
As:sp3
end projections
num_wann 12
iprint 2
dis_win_min -1000.000
dis_win_max 32.000
dis_froz_min -1000.000
dis_froz_max 15.000
num_iter 5000
dis_num_iter = 2000
-----------------------------------------------

In both cases the final interpolated bandstructures are quite similar and look as expected.

After your last email I realised that the disentanglement procedure was not converging earlier, so I had
to tweak that to get convergence.

O_D= 0.2950029 O_OD= 5.2506663 O_TOT= 21.4623126 <-- SPRD 8 Wann
5000 -0.355E-14 0.0000000600 21.4623126405 94.39 <-- CONV 8 Wann


O_D= 0.0847029 O_OD= 4.4732052 O_TOT= 23.2620467 <-- SPRD 12 Wann
5000 -0.320E-13 0.0000010973 23.2620466887 645.97 <-- CONV 12 Wann



Nandan.

[sponce]

Dear Nandan,

You need to plot the decays on a log scale (on the y-axis). Also dont use line but just points.

You should get that for large distance (large R), the decay should be 10^-5 or 10^-6.

For the plot you show, I cannot see that since it is a linear scale.

Also what is the spread on each atoms ?
The total spread seems quite small so I guess it is fine.
Make sure that the different WF center follow the symmetry you expect. The spread for symmetry equivalent WF should be the same.

Best wishes,
Samuel

[Nandan]

Hi Samuel,

I have done the calculation again from scratch and now the decay files are plotted with log scale.
They seem well converged. The electronic self-energy seems well converged up to a fine
grid of 80x80x80 q-points. I think the Wannier functions are pretty good and convergences in
the Wannier basis are reasonable.

https://drive.google.com/open?id=1yX1a1 ... rpqQjUlPO-

Again comparing two basis sets of 8 and 12 Wannier functions.

The spread for 8 Wann functions:
Final State
WF centre and spread 1 ( 0.242355, 0.242355, 0.242355 ) 3.44113417
WF centre and spread 2 ( 0.242355, -0.242355, -0.242355 ) 3.44113419
WF centre and spread 3 ( -0.242355, 0.242355, -0.242355 ) 3.44113416
WF centre and spread 4 ( -0.242355, -0.242355, 0.242355 ) 3.44113416
WF centre and spread 5 ( 1.780635, 1.780635, 1.780635 ) 2.19372387
WF centre and spread 6 ( 1.780635, 0.994899, 0.994899 ) 2.19372369
WF centre and spread 7 ( 0.994899, 1.780635, 0.994899 ) 2.19372370
WF centre and spread 8 ( 0.994899, 0.994899, 1.780635 ) 2.19372370
Sum of centres and spreads ( 5.551069, 5.551069, 5.551069 ) 22.53943164

Spreads (Ang^2) Omega I = 16.526741917
================ Omega D = 0.326531509
Omega OD = 5.686157188
Final Spread (Ang^2) Omega Total = 22.539430615


=======================
Spread for 12 Wann function
=======================
Final State
WF centre and spread 1 ( 0.340648, 0.411731, 0.337860 ) 1.86326535
WF centre and spread 2 ( 0.440946, -0.432722, -0.301533 ) 1.84828970
WF centre and spread 3 ( -0.397369, 0.463047, -0.400340 ) 1.88171063
WF centre and spread 4 ( -0.302615, -0.432624, 0.442954 ) 1.84823423
WF centre and spread 5 ( 1.596391, 1.847196, 1.593193 ) 1.63024321
WF centre and spread 6 ( 2.021258, 1.104852, 1.327255 ) 1.81328200
WF centre and spread 7 ( 1.406188, 1.451997, 0.644746 ) 1.67653003
WF centre and spread 8 ( 0.644887, 1.447845, 1.405479 ) 1.66879675
WF centre and spread 9 ( -1.351863, 2.383774, 1.423447 ) 2.38704183
WF centre and spread 10 ( -1.396343, 0.340785, 1.377987 ) 2.50362522
WF centre and spread 11 ( -1.444192, 3.885355, 2.027163 ) 1.81828845
WF centre and spread 12 ( -2.942191, 2.794267, 2.614487 ) 2.63662592
Sum of centres and spreads ( -1.384255, 15.265503, 12.492698 ) 23.57593332

Spreads (Ang^2) Omega I = 18.880768911
================ Omega D = 0.058477011
Omega OD = 4.636684015
Final Spread (Ang^2) Omega Total = 23.575929937
Final State
WF centre and spread 1 ( 0.340648, 0.411731, 0.337860 ) 1.86326535
WF centre and spread 2 ( 0.440946, -0.432722, -0.301533 ) 1.84828970
WF centre and spread 3 ( -0.397369, 0.463047, -0.400340 ) 1.88171063
WF centre and spread 4 ( -0.302615, -0.432624, 0.442954 ) 1.84823423
WF centre and spread 5 ( 1.596391, 1.847196, 1.593193 ) 1.63024321
WF centre and spread 6 ( 2.021258, 1.104852, 1.327255 ) 1.81328200
WF centre and spread 7 ( 1.406188, 1.451997, 0.644746 ) 1.67653003
WF centre and spread 8 ( 0.644887, 1.447845, 1.405479 ) 1.66879675
WF centre and spread 9 ( -1.351863, 2.383774, 1.423447 ) 2.38704183
WF centre and spread 10 ( -1.396343, 0.340785, 1.377987 ) 2.50362522
WF centre and spread 11 ( -1.444192, 3.885355, 2.027163 ) 1.81828845
WF centre and spread 12 ( -2.942191, 2.794267, 2.614487 ) 2.63662592
Sum of centres and spreads ( -1.384255, 15.265503, 12.492698 ) 23.57593332

Spreads (Ang^2) Omega I = 18.880768911
================ Omega D = 0.058477011
Omega OD = 4.636684015
Final Spread (Ang^2) Omega Total = 23.575929937

Thanks and regards,

Nandan.

[Vahid]

I have been trying to reproduce Fig.1 of the PRB paper using 700x700x700 k-grid and 100x100x100 q-grid with little success.

Looking at your input file, you are using a dense k-segment but Fig. 1 is done with a dense k-grid. Perhaps if you ran on a 200x200x200 k-grid, instead of a line segment, you get better agreement.

Vahid

Vahid Askarpour
Department of Physics and Atmospheric Science
Dalhousie University,
Halifax, NS, Canada

[Nandan]

Vahid,

Thanks for the reply.

You are right that I have used only a few points (~500) along the LGX high symmetry path to observe the Im(Sigma). I am first
trying to optimize the integration variable defined by the q-grid, and so I am varying that from 50x50x50 to just under 100x100x100.

As long as the grid size is below 100x100x100; for example 98x98x98, I get reasonable convergence. The magnitude of Im(Sigma) correpsonds
to the one in the PRB paper. As soon as the find q-grid is increased to 100x100x100, something goes wrong and I get very large values (few order
of magnitude higher) for Im(Sigma). See the figure in the link. The density of k-points between 6.1-6.3 is small, and corresponds to the Gamma minima.

https://drive.google.com/open?id=1-Z5uA ... qumjg4G0Nl

For a denser k-path along LGX I get fairly good agreement as long as fine q-mesh is under 100x100x100. This is what I am puzzled about.

Nandan.

[Vahid]

Nandan,

My tests with 700x700x700 k-grid and 100x100x100 q-grid finish successfully. I get the right magnitude for scattering rates/self-energy but instead of a nice thin line below 0.3eV (Gamma valley), I get scattered data.

It is likely that for larger q-grids, either your openmpi or fortran compiler is leaky. I have compiled my version with gcc/7.3.0, openmpi/3.1.0 and mkl/2018.2. I have had lots of problems with some earlier openmpi versions.

Vahid

[Nandan]

Hi Vahid,

I would try a lower k-grid and denser q-grid instead. If you look at the Fig 5. of Comm. Phys. Comm. Volume 209, December 2016, Pages 116-133 (the EPW paper),
as the q-grid becomes denser the scatter reduces.

Thanks for the clue about the compilation issue. I have installed QE-6.2.1 this week with Intel 17.0 and OpenMPI-3.0.0 and running GaAs for a 100x100x100 grid.

Nandan.

[vinpandey]

Dear users,
can you tell me about convergence criteria , when we do the different fine qgrid calculation?
as you have said earlier, for a polar material you may need 300x300x300 qgrid , which is actually very expensive to be calculated.
Actually, I am working on PbI2, where I use qgrid 300x300x20(1.8 million qgrid). and epw is progressing with 5000 qgrid per hour, it says, it will go for 15 days approx. I am doing calculation on 112 processor, that too when I am not using enough qgrid along z axis.
can you tell me if there is any way to speed up the calculation or I can relax the convergence criteria a bit like 2 % or 3% change in (Im linewidth) with change in different higher qgrid.