Yambo vs Gaussian

[Daniele Varsano]

Dear Martin,
as I just told you, BSE without screening reduces to the Time-Dependent Hartree Fock, I invite you to demonstrate it.
Reasons for discrepancies could be of any nature, and you should investigate them, for instances did you compare your GW gap with the literature?
May be you have a too high quasiparticle gap.
If you want you can post your inputs/outputs/reports and I can have a look if there is something odd at first glance.
Best,
Daniele

[martinspenke]

Dear Daniele,

Attached are my inputs and results without using Cut-Off since it worsened the results.
I compared gw band gap to the literature. Experimental band gap of isolated water molecule is 12.5 eV , and i have 12.49 eV. (states : 4 - 5) (converged result)
I am not interested in TDHF because it gives any way not the desired results as you can see in the latest paper to which i refereed (page 39 table 3).

Bests
Martin

[Daniele Varsano]

Dear Martin,
Unfortunately I'm not expert with water,
In your input the only think that look me strange is the number of G vecotrs in GW, 500 Gvec with respect the total charge looks me not enough, you can try to consider much more G vectors, possibly in PP model. In the PRL of Garbuio they used an order of magnitude more (reported in the bibliography).
Anyway, your GW calculation is in agreement with with previous calculation in isolated molecule:
P. H. Hahn et al., Phys. Rev. Lett. 94, 037404 (2005).
http://users.physik.fu-berlin.de/~ag-gr ... t_Talk.pdf

Even if I remembered that there were problem in GW, in particular with G0W0 for water:
Phys. Rev. B 81, 085103
http://www-ipcms.u-strasbg.fr/IMG/pdf/rinke.pdf
may be could instructive to look at DeltaSCF results.

In any case, assuming that GW is correct you have a rather big discrepancy with respect the results of Bechstedt:
did you check if the first peak is mainly homo-lumo? (you can see by diagonalizing the BSE, as may be even if the gap is converged, higher level could not).
As it looks that electrostatic effects are very important passing from solid to molecules, the only think I can suggest you is to increase a lot your supercell and add the coulomb cutoff, in the report you posted I could not see the adopted geometry.
Next you can also consider to add bands in BSE. Moreover note that the TD correspond to set BSmod="resonant", but this will not change nothing.
Unfortunately in the paper of bechstedt technical parameter are not reported, may you be can write to the corresponding author.

I do not if it helps, but I'm not expert on that and a quick look to the literature gives me the impression that it is rather problematic.

Daniele

[martinspenke]

Dear Daniele,

Thank you very much.
But unfortunately this problem holds for other molecules, too.
I also tried to get excitation energies of some organic molecules, like Furan, Naphthalene, etc (my first post), too.
But the story is the same, excitation energies are blue shifted and totally wrong as i compared it to the gaussian results.
So this here is not a only water specific problem. I have to contact people providing excitation energies of molecules with bse in their papers.

Any way thanks a lot for your time and helps.
Best wishes
Martin

[Daniele Varsano]

Dear Martin,
I did several calculations for molecules in BSE (I posted a couple of papers), actually I implemented the method to isolate the molecule.
In my experience converge results for molecules in plane wave is a quite painful task, and systematic tests are needed.
For instances, in the case of Furan you posted, beside the molecule was not isolated your exchange part of the BSE had very few Gvector (only 739),
and the spectrum was redshifted (few spurious dark states) and bright states around 0.3eV lower than the reference values. I noticed now that also for the water case you have few exchange Gvectors (look at the output and report! the same you have in correlation), beside the fact you put an high number in the input. Actually that should be a bug!!, I will look at it,in the meanwhile you can add the flag ALLGexx in input. Most probably setting the proper number of gvecotrs in exchange will blueshift the excitation energies.
In summary I think you need to consider bigger cells (in particular for water which is polar), isolate them properly with the coulomb cutoff respecting the condition between system size, box size and cutoff radius, and systematically converge the parameters (50 bands in BSE most probably are not enough, adding bands should redshift your results). It is important that the cutoff is used for calculating the screening and the rest of steps. This most probably will not give you results matching experiments and QC calculations, but will give you meaningful results and the discrepancies have to be understood in terms of the approximation of the theory. Moreover I suggest you to use the PP approximations, even if not totally justified, it usually provide reasonable results and I'm not confident on the convergence of the real-axis with respect bands and energy steps, and as the calculation explode adding energy steps, it is very hard to converge it.
Another suggestion, do not use bands runlevel in qe runs, but use nscf.
Unfortunately this kind of calculations are not plug-and-play and a good dose of patience is needed.
Hope it helps,
Daniele

[martinspenke]

Dear Daniele,

. I noticed now that also for the water case you have few exchange Gvectors (look at the output and report! the same you have in correlation), beside the fact you put an high number in the input. Actually that should be a bug!!, I will look at it,in the meanwhile you can add the flag ALLGexx in input. Most probably setting the proper number of gvecotrs in exchange will blueshift the excitation energies.

Oh my god, i see this.

many many thanks, i am going now to do things you mentioned.
Any way i get good converged qp energies using real axis.
I raised again my NGsBlkXd in GW calculation without cut-off from 500 to 1037, and noticed only a very small decrease of 0.09 eV in the band gap.

I will be totally happy if i can get excitation energies in the range of 0.1 to 0.3 eV from experiments for low lying states like that table on page 39 for water.

I contacted Patrick Rinke for gw band gap of water , telling me " The best plane wave G0W0@PBE value for water is 11.68 eV" but i have to check this again.

Bests
Martin

[Davide Sangalli]

Dear Martin,
it is not exactly a bug ...

In any case you can avoid this problem uncommenting the variable "ALLGexx" in your input file.
Maybe we will improve this part setting the "ALLGexx related logical" to true by default

Best regards,
Davide

[martinspenke]

Dear Davide,

it is not exactly a bug ...
In any case you can avoid this problem uncommenting the variable "ALLGexx" in your input file.

Yes i agree, this is not a bug. By the way i get this warning :

<01s> [WARNING]Exchange FFT size is too big. RL vectors reduced to 4015

probably because of memory reasons.

I noticed now that also for the water case you have few exchange Gvectors (look at the output and report! the same you have in correlation), beside the fact you put an high number in the input

I can not have more because i do not have memory available any more. I only have 32 Gb on my system.

I will look at it,in the meanwhile you can add the flag ALLGexx in input. Most probably setting the proper number of gvecotrs in exchange will blueshift the excitation energies.

Unfortunately i need a red shift not a blue shift any more.

To make a isolated molecule using coulomb cut-off results in slightly wrong qp energies and hence wrong gw gap which leads to a wrong shifted bse spectrum.
I do not understand fully the usefulness of this coulomb cut-off beside having the possibility of reducing computation time using a smaller simulation box than usual,
if it generates uncontrollable errors in qp energies which falsify the final spectrum !!!???!!!!

(50 bands in BSE most probably are not enough, adding bands should redshift your results). It is important that the cutoff is used for calculating the screening and the rest of steps.

I also tried it with 120 band in BSE but no change, still the same. I ALSO have done all of these calculation stuffs with coulomb cut-off in every step using sphere xyz and a radius of 12.5 a.u (half of my box) too, but as i said above, this technique produces wrong unsatisfactory results.

Another suggestion, do not use bands runlevel in qe runs, but use nscf

I used a scf calculation in QE but i think the difference between scf and nscf is of only 0.1 eV. No big deal.


Best Regards
Martin

[Daniele Varsano]

Dear Martin,

probably because of memory reasons.

Memory issues does not enter here at all. Use the flags me and Davide suggested and you can raise the values of G vectors in exchange. What you are using is clearly out of convergence.

Unfortunately i need a red shift not a blue shift any more.

You cannot force the calculation to match what you need. Unconverged calculations can match what you aim, but they are meaningless, and the computational spectroscopy losses his sense.

To make a isolated molecule using coulomb cut-off results in slightly wrong qp energies and hence wrong gw gap which leads to a wrong shifted bse spectrum. I do not understand fully the usefulness of this coulomb cut-off beside having the possibility of reducing computation time using a smaller simulation box than usual, if it generates uncontrollable errors in qp energies which falsify the final spectrum !!!???!!!!

QP energy are charged excitations, convergence with the volume of the supercell is terribly slow, and converge the results with respect to the volume became unfaisable. This is documented in various publications. Again, uncontrollable error respect to what? experiments? please see my comment above.

Convergences:

I raised again my NGsBlkXd in GW calculation without cut-off from 500 to 1037, and noticed only a very small decrease of 0.09 eV in the band gap. [/quote
This is not a convergence test, as it is based on only two calculations. The convergence could be simply slow, you need to look at a trend, 0.1eV each 500 Gblk can bring you very far away, I'm not saying this is the case as I cannot know, but this has to be checked. Moreover, you continue to use real-frequency methods, did you check the convergence with respect the number of steps you are performing integrations?

I used a scf calculation in QE but i think the difference between scf and nscf is of only 0.1 eV. No big deal.

GW/BSE is a perturbation theory, if you start your calculations with not well converged ground state you can arrive to very different final results. The procedure on how to start a Many Body Perturbation Theory calculation is explained in the documentation. If you do not want to follow it, of course feel free to do as you prefer.

Best,
Daniele

[Daniele Varsano]

Dear Martin,
I can see there is an user that are doing calculations similar you were doing on water and it looks got results similar to the paper by Bechstedt:
If you are still interested in it, you may contact him to exchange some info:
This is the post:
viewtopic.php?f=15&t=849#p3758


Best,
Daniele