from readmat : error # 1 wrong qpoint

[dinesh169]

Hello Raxona and Carla,

Thank you very much for quick reply. But I wonder why does it not even read the very first q-points. Because there <QPOINT_NUMBER type="integer" size="1"> is 1.


Best wishes

[dinesh169]

Hallo Raxona and Carla,

Thank you very much for your replies. That wrong q point was the problem in the pattern#.xml files. I tested some calculations. It worked out.

Have a good day.

Best wishes,
Dinesh

[dinesh169]

Hello Everyone,

I am calcualating slef energy for bulk mos2. I have a question about the scaling of the code.
I used 64 processor and the job did not even finish after 17 days.
Here is the input file:

--
&inputepw
prefix = 'mos2'
amass(1) = 32.065,
amass(2) = 95.94,
outdir = './'



iverbosity = 0

elph = .true.
epbwrite = .true.
epbread = .false.

epwwrite = .true.
epwread = .false.

nbndsub = 30
nbndskip = 0

wannierize = .true.
num_iter = 10000
iprint = 2
dis_win_max = 19
dis_froz_max= 19
proj(1) = 'random'
kmaps = .false.


elinterp = .true.
phinterp = .true.

tshuffle2 = .true.
tphases = .false.

elecselfen = .true.
phonselfen = .false.
a2f = .false.

parallel_k = .true.
parallel_q = .false.

fsthick = 136056981 ! eV
eptemp = 300 ! K (same as PRB 76, 165108)
degaussw = 0.1 ! eV
max_memlt = 4.0



dvscf_dir = './save'
filukk = './mos2.ukk'

nkf1 = 60
nkf2 = 60
nkf3 = 20
nqf1 = 60
nqf2 = 60
nqf3 = 20

nk1 = 12
nk2 = 12
nk3 = 4

nq1 = 6
nq2 = 6
nq3 = 2
/
14 cartesian
0.0000000 0.0000000 0.0000000 0.0277778
0.0000000 0.0000000 -0.1301886 0.0277778
0.0000000 0.1926310 0.0000000 0.1666667
0.0000000 0.1926310 -0.1301886 0.1666667
0.0000000 0.3852620 0.0000000 0.1666667
0.0000000 0.3852620 -0.1301886 0.1666667
0.0000000 -0.5778930 0.0000000 0.0833333
0.0000000 -0.5778930 -0.1301886 0.0833333
0.1668233 0.2889465 0.0000000 0.1666667
0.1668233 0.2889465 -0.1301886 0.1666667
0.1668233 0.4815775 0.0000000 0.3333333
0.1668233 0.4815775 -0.1301886 0.3333333
0.3336467 0.5778930 0.0000000 0.0555556
0.3336467 0.5778930 -0.1301886 0.0555556

I read on forum that max no of processor that can be used are nk1*nk2*nk3. Can you please tell me how does calculations scales with number of processors? Does num_bands really effect the scaling significantly?

Best wishes,
Dinesh

[sponce]

Dear Dinesh,

A scalability plot is available in our publication for the code:

http://www.sciencedirect.com/science/ar ... 5516302260

However the scalability has been improved. I want to produce a new graph for that. It should scale quite well up to ~250 cores.

The max number of processor that can be used are nk1*nk2*nk3 FOR THE FIRST PART only !

In the new EPW 4.1 version of the code, you can use a maximum of nkf1*nkf2*nkf3 cores for a restart calculation. This means that you first need to have the XX.epmatwp1 file correctly written.
You can then restart the calculation from there with epwread = true.
Note that you also need to set kmaps to true.

For your other question, num_bands scales as square ! So yes, it does affect significantly.

With the new version you need to remove some depreciated input variables. You can read more at the bottom of http://epw.org.uk/Main/About

Cheers,

Samuel

[dinesh169]

Hello Everyone,

I tried to run the above calculations twice with 100 processors and it ran for almost one month and then died without finishing.
Then I thought of skipping some lowest valence bands to get the calculations done. For this I tried graphene as example, when I use num_wann = num_bands or num_wann > num_bands (by defining some energy windows) everything works well I mean DFT and wannier band structure looks identical. But on using nbndskip = INTEGER, then the wannier bandstructure (G-K-M-G) looks very different in the M-K region all the bands are shifted from N to N-1 and again. Conduction band enters in valence band in M-K region with some energy eigen values in forbidden region which rules out the possibility of wrong sorting. With nbndskip I define frozen and wannier window such that it exclude the lowest bands.
Here is the wannier part of the calculations

nbndsub = 9
nbndskip = 1
kmaps = .false.

wannierize = .true.
num_iter = 2000
iprint = 2
proj(1) = 'random'
dis_win_max = 12
dis_win_min = -24.604999
dis_froz_max= 9.623
dis_froz_min= -14.6049999


wdata(1) = 'bands_plot = true'
wdata(2) = 'begin kpoint_path'
wdata(3) = 'G 0.000000000 0.000000000 0.000000000 K 0.333333333 0.333333333 0.000000000'
wdata(4) = 'K 0.333333333 0.333333333 0.000000000 M 0.500000000 0.000000000 0.000000000'
wdata(5) = 'M 0.500000000 0.000000000 0.000000000 G 0.000000000 0.000000000 0.000000000'
wdata(6) = 'end kpoint_path'
wdata(7) = 'num_print_cycles = 50'
wdata(8) = 'conv_window = 2'
wdata(9) = 'conv_tol = 1.0E-10'
wdata(10)= 'dis_conv_tol = 1.0E-10'
wdata(11)= 'dis_mix_ratio = 0.5'
wdata(12)= 'dis_num_iter = 8000'

I used 12x12 and 18x18 grid to check the interpolation, but the results were same. Can you please suggest something?

Best wishes,

[sponce]

Hello,

What is the output of the Wannier calculation?

Maybe you can try to use a better starting projection. Usually 'random' is quite bad.

Also, excluding 1 band will not help a lot.

However you will be a major speed improvement going from nbndsub = 30 (your previous post) to nbndsub = 10.

So if I understand correctly, there was 20 conduction bands before?

Best,

Samuel

[dinesh169]

Hello Dr. Samuel,

Thank you very much for your reply.
I solved the problem of band interpolation, the choice of energy window was crucial.
I know exclusion of 1 band wont be helpful, as I mentioned before I was trying the calculations
on graphene just to make sure what I do is correct.
As far as projections is concerned, by having a look on the hybridization state of molecule one can
specify those orbitals as initial guess. What if it does not cover the number of bands we want to include in the calculation??
Is there any generalised way to define the projections? What I do is, start with random and use 10,000 iterations and spread
is around 1-2 A for every orbital and check band structure.

Now I am facing another error when trying to use "ephwrite = .true." in the input file.
Error is as follows:
forrtl: severe (71): integer divide by zero


Here is the input file
1;2305;0c--
&inputepw
prefix = 'gr'
amass(1) = 12.01078
outdir = './'

iverbosity = 0

elph = .true.
epbwrite = .true.
epbread = .false.
ephwrite = .true.

epwwrite = .true.
epwread = .false.

nbndsub = 8
nbndskip = 0

wannierize = .true.
num_iter = 500
iprint = 2
proj(1)= 'C :random'
! proj(2)= 'f=0.51701, 0.11939, 0.98543 :s'
! proj(3)= 'f=0.75879, 0.80509, 0.63419 :s'
! proj(4)= 'f=0.08238, 0.57572, 0.33325 :s'
! proj(5)= 'f=0.59309, 0.48622, 0.19279 :s'
! proj(6)= 'f=0.61222, 0.70488, 0.87394 :s'



! dis_win_max = 20
! dis_win_min = -17.0
! dis_froz_max= 15
! dis_froz_min= -17


wdata(1) = 'bands_plot = true'
wdata(2) = 'begin kpoint_path'
wdata(3) = 'G 0.000000000 0.000000000 0.000000000 K 0.333333333 0.333333333 0.000000000'
wdata(4) = 'K 0.333333333 0.333333333 0.000000000 M 0.500000000 0.000000000 0.000000000'
wdata(5) = 'M 0.500000000 0.000000000 0.000000000 G 0.000000000 0.000000000 0.000000000'
wdata(6) = 'end kpoint_path'
wdata(7) = 'num_print_cycles = 50'
wdata(8) = 'conv_window = 2'
wdata(9) = 'conv_tol = 1.0E-10'
wdata(10)= 'dis_conv_tol = 1.0E-10'
wdata(11)= 'dis_mix_ratio = 0.5'
wdata(12)= 'dis_num_iter = 8000'

elinterp = .true.
phinterp = .true.

tshuffle2 = .true.
tphases = .false.

elecselfen = .false.
phonselfen = .false.
a2f = .false.

parallel_k = .true.
parallel_q = .false.
filqf = './qpoints.dat'
filkf = './qpoints.dat'
degaussw = 0.025

dvscf_dir = './save'
filukk = './gr.ukk'
! nkf1 = 12
! nkf2 = 12
! nkf3 = 1

! nqf1 = 6
! nqf2 = 6
! nqf3 = 1

nk1 = 12
nk2 = 12
nk3 = 1

nq1 = 6
nq2 = 6
nq3 = 1
/
7 cartesian
0.0000000 0.0000000 0.0000000 0.0555556
0.0000000 0.1924501 0.0000000 0.3333333
0.0000000 0.3849002 0.0000000 0.3333333
0.0000000 -0.5773503 0.0000000 0.1666667
0.1666667 0.2886751 0.0000000 0.3333333
0.1666667 0.4811252 0.0000000 0.6666667
0.3333333 0.5773503 0.0000000 0.1111111

Thanks in advance

Best wishes,
Dinesh

[sponce]

Dear Dinesh,

Could you update your version of EPW to the latest 4.1 (from QE 6.0) please?

Note that those input variables have been removed:
elinterp = .true.
phinterp = .true.

tshuffle2 = .true.
tphases = .false.

Hopefully this will solve your problem.

In the file you have sent, why did you comment the dis_win_max etc if the windows are crucial?
Since you have wannierize = .true., you will do the Wannierization.

Best,

Samuel

[dinesh169]

Hello Samuel,

Okay, I will check with the updated version.

I comment the windows because now I want to check the phonon interpolation. wannier_functions= num_bands, so that in case I am spanning whole energy range.

Best wishes,
Dinesh

[sponce]

Dear Dinesh,

The phonon interpolation is done via Fourier interpolation (like when you use q2r.x and matdyn.x) and not a Wannier interpolation.

You can therefore safely use the same reduced range of band with the windows for the phonon interpolation.

You can test that it will give the same interpolated phonon frequencies regardless of the number of electronic bands included.

Best,

Samuel