Plasmon pole approximation stops at 10%

[ivasan]

Dear all,

I am finding a strange output in the log file when running GW in the plasmon pole approximation. The strange thing is the following:

<08h-28m-50s> [06] Dyson equation: Newton solver
<08h-28m-50s> [06.01] G0W0 : the P(lasmon) P(ole) A(pproximation)
<08h-28m-50s> [M 0.808 Gb] Alloc WF (0.804)
<08h-28m-50s> [FFT-SC] Mesh size: 27 27 27
<08h-28m-50s> [M 1.647 Gb] Alloc wf_disk (0.838)
<08h-28m-50s> [WF-SC loader] Wfs (re)loading | | [000%] --(E) --(X)
<08h-28m-57s> [WF-SC loader] Wfs (re)loading |# | [005%] 06s(E) 02m-13s(X)
<08h-29m-04s> [WF-SC loader] Wfs (re)loading |##### | [025%] 13s(E) 55s(X)
<08h-29m-11s> [WF-SC loader] Wfs (re)loading |########## | [050%] 21s(E) 42s(X)
<08h-29m-19s> [WF-SC loader] Wfs (re)loading |############### | [075%] 28s(E) 38s(X)
<08h-29m-19s> [WF-SC loader] Wfs (re)loading |####################| [100%] 29s(E) 29s(X)
<08h-29m-19s> [M 0.808 Gb] Free wf_disk (0.838)
<08h-29m-19s> G0W0 PPA | | [000%] --(E) --(X)
<10h-09m-54s> P02: G0W0 PPA |# | [005%] 03h-39m-21s(E) 03d-01h-07m-04s(X)
<20h-22m-14s> P04: G0W0 PPA |# | [005%] 07h-19m-22s(E) 06d-02h-27m-20s(X)
<10h-11m-48s> P03: G0W0 PPA |# | [005%] 03h-39m-44s(E) 03d-01h-14m-50s(X)
<17h-35m-39s> P01: G0W0 PPA |## | [010%] 09h-06m-19s(E) 03d-19h-03m-13s(X)
<20h-53m-37s> [M 0.004 Gb] Free WF (0.804)
<20h-53m-37s> [06.02] QP properties and I/O
<20h-53m-37s> [07] Game Over & Game summary

The G0W0 PPA calculation stops when the 10% is reached without reporting any error, and this happens with all tests I have run. The problem is that afterwards I try to obtain the photoemission spectra, the results seems quite strange.

I will described in the following the case I am study and what are the strange results.

YAMBO VERSION: 2.3.4.r.855
COMPUTER: SP6@CINECA
COMPILATION OPTIONS: I compiled yambo myself using the following options:

export CPP=cpp
export CC=xlc_r
export F77=xlf_r
export FC=xlf90_r
export FCFLAGS='-O2 -q64 -qstrict -qarch=pwr6 -qtune=pwr6 -qmaxmem=-1 -qsuffix=f=f'

module load xlc
module load xlf

./configure --build=powerpc-ibm --enable-dp

(a very simple configuration)

SYSTEM UNDER STUDY: cubic simulation cell containing 219 silicon atoms with periodic boundary conditions (KSS file obtained with abinit).

PROCESS TOWARDS THE GW-PPA:
a2y -S -F file.kss
yambo -F y01.i.in (input file obtained with yambo -i -V RL -F y01.i.in)
yambo -F y02.gw.tbands.in (input file obtained with yambo -p p -g n -F y02.gw.tbands.in)
yambo -F y02.gw.rpa.wqp.tbands.in (input file obtained with yambo -o c -V qp -F y02.gw.rpa.wqp.tbands.in)

I have attached all the input files (y0*.in), the resulting log and report files, and the scripts used in CINECA (job.*.cineca) to submit the job. I cannot attach the kss file since it is a 3.4 GB file.

files.tar.gz

WHICH ARE THE STRANGE RESULTS?
Initially I have done a convergence test in order to find the minimum number of empty bands to use.
In the Fig1.jpg it is represented Img( epsilon ) evaluated with ‘-o c’ and corrected with the QP energies (see y02.gw.rpa.wqp.tbands.cmd) for different numbers of empty bands. For a faster calculation I made NGsBlkXp=1. It can be seen that 638 bands are enough for a qualitative description of the spectra, while at least 738 give the convergence.

Fig1.jpg

The next step is converging NGsBlkXp. For this purpose I run calculations with 638 and 738 empty bands, and NGsBlkXp = 500 and 600. The results are shown in Fig2.jpg. Then, the convergence is lost when NGsBlkXp is different from 1, which might be due to the fact that the G0W0 PPA is not correctly computed for my particular system.

Fig2.jpg

I will appreciate any help with this issue.

Thanks in advance for your answer and for your time.

Kind regards,

Iván

NOTE: the attached files correspond to the case of 638 bands with NGsBlkXp = 600.

[ivasan]

Dear yambo developers,

In order to understand what was going on with the calculations, I printed the first elements of the dielectric matrix. It seems that we have not reach the convergence on the screening. This explains why the single particle spectra behaves as it was shown on the previous post.

Nevertheless, there is still open the question about the calculations in plasmon pole approximation: they always stop when the 10% is reached. Is there any way to check if the calculation is correctly finished by printing a variable or anything?

Thanks in advance for your time and answer.

Regards,

Iván

[myrta gruning]

Hallo Ivan

The problem of the 10% may be simply due to the fact that the next steps, for some reason are not written in the log file. It will be indeed strange that the calculation will continue without any error.
If you are linking with netcdf you can use ncdump command to inspect the SAVE/ndb.pp that should contain the plasmon pole part.

Anyway be careful, it seems to me you are mixing things a bit.
In fact if there is a problem with the calculation of the plasmon pole part this will influence just the GW corrections and NOT the calculation of the Chi (your calculation is at RPA level).

Best,
m

[ivasan]

Dear Myrta,

Thanks for your answer. Indeed I am quite new in these issues and I still have to assimilate better some concepts.

In my calculations I am including the GW corrections to the energy levels when evaluating Chi by using
-o c -V qp
and then by specifying
XfnQPdb= "E < ./SAVE/db.QP"
in the input file.

Regarding the Plasmon Pole (PP) approximation, I am not linking to NETCDF. I just did a simple compilation at CINECA (the details are in the first post), and the file containing the information for the PP calculation is SAVE/db.pp. Is there any other way to check this file?

Thanks

Iván

[myrta gruning]

Thanks for your answer. Indeed I am quite new in these issues and I still have to assimilate better some concepts.

In my calculations I am including the GW corrections to the energy levels when evaluating Chi by using
-o c -V qp
and then by specifying
XfnQPdb= "E < ./SAVE/db.QP"
in the input file.

Yes you are right in the sense that the plasmon pole will influence the corrections, then the spectra. I was a bit confused by the fact that you are checking the converge of the plasmon pole parameters from the optical spectra instead of looking at the behavior of the qp corrections in o.qp...

Regarding the Plasmon Pole (PP) approximation, I am not linking to NETCDF. I just did a simple compilation at CINECA (the details are in the first post), and the file containing the information for the PP calculation is SAVE/db.pp. Is there any other way to check this file?

I looked better at your log. Also in this case I was a bit confused in my first answer (sorry I am just back from holidays... it needs some days to get back in the routine )
The plasmon pole is computed correctly since the X calculation is finished. The calculation that seems not to be finished is the one for the QP corrections... Following the log it should be 3 days long, but it stops after few hours (about 12h). In this case you can look at o.qp, if all corrections are there and are "reasonable".
I see if I can reproduce this behaviour and see whether it is a problem with logging (hopefully) or something else.

Best
m

[myrta gruning]

Hallo again,

I see if I can reproduce this behaviour and see whether it is a problem with logging (hopefully) or something else.

Indeed the problem can be reproduced in any GW calculations.
The problem is related to the logging, the expected time for calculation is not calculated correctly, is far too large, that is why it seems that the calculation stopped at 10%. In fact yambo correctly complete the calculation for the gw corrections.
I fixed the problem. I will test the fix and then soon apply it to the source in the repo.

Best
m