Gap convergence with vacuum space between layers

[Daniele Varsano]

Dear Fabio,
what I can see from the report is that the first unoccupied bands for some k points has a change of ~.1eV already at DFT level, which is quite disturbing. Beside that i suggest you to look at first at VBM and CBM only:
%QPkrange # [GW] QP generalized Kpoint/Band indices
| 7| 7| 4| 5|
%

repeating your calculation (exchange only) for different vacuum space by setting:
EXXRLvcs to the MaxGvector for each case, here in the report for the case c20:

Code: Select all

  G-vectors             [RL space]:  38279

and do not lower the FFTGvecs, keep it to the maximum (actually you are using PBE and it needs the G vectors of all the charge to be calculated).
Next see if the HF energies for these two bands have a converging behaviour wrt the vacuum.
Actually yes, the problem seems to come from the non-local part (HF), but try to do a clean test ie with all the Gvectors in EXXRLvcs and FFTGvecs (if you do not specify it should be the maximum).

Next you can try to activate the RIM (add -r in input), in order to see if the problem is in the integrals in the BZ when it is largely anysotropic.
You can set something like:

Code: Select all

RandQpts=1000000                   # [RIM] Number of random q-points in the BZ
RandGvec= 1            RL    # [RIM] Coulomb interaction RS components

for the different case and see if it solves the problem.

Best,
Daniele

[Fabiof]

Dear Daniele,

When applying the RIM method, do i need to set the truncation variables? Or can i just use

RandQpts=1000000 # [RIM] Number of random q-points in the BZ
RandGvec= 1 RL # [RIM] Coulomb interaction RS components

Fabio

[Daniele Varsano]

Dear Fabio,
I was suggesting to set the RIM without andy truncation, so leaving the bare coulomb potential in order to see if the problem of the departure from a converging
behaviour is related to the BZ integration.

Best,
Daniele

[Fabiof]

Dear Daniele,

In the latest calculations i used always the EXXRLvcs to the MaxGvector and never changed the FFTGvecs.

Using the RIM method i got the following results

c bohr gap eV
20 4.43172
30 4.5226
45 4.79309
60 5.17023
70 5.44001
80 5.74752


I saw some papers that using Gap vs 1/vacuum distance you can extrapolate the gap at infinite.

But in this case (RIM) and without RIM i can't do that extrapolation because when i plot Gap vs 1/vacuum it has a exponential behavior.

Fabio

[Daniele Varsano]

Dera Fabio,
sorry I'm bit confused, what is here the difference with respect the values you report in your first post?
Only the RIM?
In your first post you had a converging behaviour up to 100au.

Daniele

[Fabiof]

Dear Daniele,

Sorry i am not being very clear in this topic and it is my fault. I will recalculate from 20 to 100 without and with RIM and post the results for a clear comparison.

Fabio

[Fabiof]

Continuing the last post.

I recalculated the results with and without RIM with the variables

XXRLvcs = MaxGvector for each vacuum distance
NGsBlkXp = 4Ry
BndsRnXp =100
GbndRnge =100


Results without RIM:

vacuum Bohr vs Gap eV

20 4.37373
30 4.32706
40 4.48611
60 5.03015
70 5.32016
80 5.64496
100 6.3249
150 8.06742


results with RIM
RandQpts=1000000 # [RIM] Number of random q-points in the BZ
RandGvec= 1 RL # [RIM] Coulomb interaction RS component


20 4.39905
30 4.3995
40 4.54707
60 5.03839
70 5.30683
80 5.61556
100 6.26765
150 7.96562


I don't see improvements in the gap convergence. What do you think?


Fabio

[Daniele Varsano]

Dear Fabio,
ok so it is not a problem of integration. Anyway:
1) Why so much difference from the first results you posted, where a converging behaviour was observed up to a certain distance.
2)

Code: Select all

 BndsRnXp =100 , GbndRnge =100 

It does not make much sense to maintain the same number of bands for each calculations, if you look the maximum energy you reach with 100 bands you will see bug differences with 20 and 100 au of vacuum, so the level of convergence of the calculations are not consistent. You should change the number of bands in order to have similar energy included, which means that for larger box they will be too many.
3) Check also if the DFT gap does not vary in the different cells.
4) Try to make consistent tests, and then as said before try to see if the problem comes from the exchange or the correlation part.

According to the first values you posted days ago, it looked that it was converging till some value and next the calculation with the larger boxes were totally out-of-convergence, but it was just an impression. Now with these new calculations I do not know what to think, but again they are strongly biased from the fact you kept the number of band fixed.

Best,
Daniele

[Fabiof]

Dear Daniele,

Thanks for the answer.

1) Yes it is really odd the results in first post are different. The only difference is in the parallel scheme, but i think that i have made some mistake. The recalculated gaps that i posted in the last post without RIM are in agreement with some calculations that i did before. So i would say that is very probable i did something wrong in the first post.

2) I have checked before what happens when increasing to 1000 bands for c = 20,30,45,70 and 80 bohr. What i saw is that the convergence for each vacuum distance is achieved for 500 or 600 bands (BndsRnXp = GbndRng). Also increasing the bands will increase the gap value. So if i let the c=20 bohr with 100 bands and c=80 bohr with 1000 or more bands, i think it is going to make the convergence worst because the gap value of c=80 with 100 bands is lower than the value with c=80 with 1000 bands. But i will recheck this.

3) There are no differences between c=30 and c=100 Bohr. c =20 Bohr is a very small distance and it makes the indirect gap a little bit lower than the other distances.

4) Ok, i will do more tests.



Thanks

Fabio

[Daniele Varsano]

Dear Fabio,
1) Well, this is strange, the results in the first post made some sense, the latter honestly it is something I've never seen before.

Code: Select all

 The only difference is in the parallel scheme

Well if you observed this, please report as it *should* not happen at all.

3) Ok this i s a good starting point.

Anyway, convergence with vacuum is very hard and for this reason the cutoff coulomb potential it helps a lot. In any case, without the cutoff you should observe a convergence even if very slow, and for some reason you are not observing.
Again I think it is useful to separate exchange and correlation in order to understand the source of the problem

Best,

Daniele