Xd damping range

[martinspenke]

Dear Developers,

I noticed that the G0W0 energies in case of molecular systems (in a large super-cell) strongly depend on the Xd damping range, if one decreases it to near zero (0.01,0.001,.0.0001.etc ...).
% DmRngeXd
0.10 | 0.10 | eV # [Xd] Damping range
%

So, my question is whether the convergence of QP energies (for ionization potential) should be checked with respect to DmRngeXd or a small damping of, let say 0.1 eV, is mandatory to have ?

Best wishes
Martin

[Davide Sangalli]

Dear Martin,
I would say the smallest possible smearing should be used in principle, i.e. sending the "eta parameter" to zero as one does theoretically.
I think "0.001" is reasonable, instead 0.1 and 0.01 are still quite "big".

Do you see any significant difference in between 0.001 and 0.0001 ?

D.

[martinspenke]

Dear Davide,

many thanks for your response.

I also think that Xd damping should be chosen as small as possible.
I get a converged G0W0 ionization potential for 0.00001 eV for Helium atom.

However, the problem is that in literature the G0W0@PBE is 23.39 eV with the WEST code (and other codes like VASP)
http://www.west-code.org/database/gw100 ... -7.php#vip
I can also nearly reproduce this number using PW+YAMBO for xd (eta parameter) = 0.1 eV, but as soon as I decrease it to 0.001, 0.0001 eV, the ionization potential is INCRESAED by 1 eV.

ALL the other computational parameters such as ECUTT, bands, dielectric cutoff, super-cell size are chosen to be converged.

What would you advice ?
I mean setting eta = 0.001 or 0.0001 eV clearly falsifies what is published so far!


Best wishes,
Martin

[Daniele Varsano]

Dear Martin,
this is a good point.
I think that convergence tests with respect the damping are not so usual. Honestly me personally I do not use to check it and I'm pretty sure I'm not the only one, and I think most of the people do not pose the attention. My suggestion is to contact West people (they use to be nice!!) and ask details of their calculations. If they agree to repeat their calculation with a smaller damping, a double check with a different code I think that will be helpful for everyone.
Best,
Daniele

[martinspenke]

Dear Daniele,

1) If having a small smearing is important (what I believe, because in an old paper by andrea marini, the QP lifetimes increased by up to 50% when decreasing the smearing)
then I am quite sure that all published ionization potential (IP) results of small molecular systems (http://pubs.acs.org/doi/abs/10.1021/acs.jctc.6b01150) are simply unconverged.


2) Another aspect which I like to share with you is that I used the sphere cut to avoid artificial screening, and noticed that IP energy oscillates with the number of G vectors!
I.e., when using a kinetic energy cut-off of 100 Ry and 140 Ry for the same super-cell size with sphere cut-off I have oscillations in my IP energy by almost 1 eV!

However, whenever I totally neglect the cut-off procedure in Yambo, my IP energy with respect to different kinetic energy cut-off energies becomes stable.

To me it seems that this sphere cut procedure somehow suffers from instabilities for large G-vectors!

Best wishes,
Martin

[Evgeniia]

Hi! I have just met the same problem, and I think that either this might be the bug in the code or I do not understand the physical meaning of this parameter. So I need some advise from admins.
As I understand this damping range just gives the width of the lorentzian for the energy peak. if it is too large, then I will see the broadening of my data. If it is small, the data will be peaked, but the integral (the square under the curve ) should be kept the same. So according to my understanding there is no "convergence" as we use this term here.

I have checked the damping range 0.1, 0.01 and 0.001 without linear increase for all other parameters kept the same. I attach the graph here. It seem for me, that something is wrong with the norm conservation, because the shape of the graph is the same for 0.01 and 0.001, but the integral is different. (for the red curve and blue curve it differs as well, but not so significantly- approximately for 15%)

Additionally I decided to check the code at some limits. I used only the gamma point and only homo and lumo orbitals to calculate BSE result. The thing is that for the damping 0.1 everything works well. But if I change the damping to 0.15 or, contrary,make it smaller than 0.1 the Haydock threshold (which is 0.02 by default) stops working, and I need to change it to 10% or higher. So, it is the question, why everything works only for the default damping range.

Unfortunately I cannot upload the comparison graph here, so i put the google drive link. Please tell me, if it breaks the forum rules.
https://drive.google.com/open?id=1hgAnD ... 18Zt4Kyf0a

Could you please comment on this?

Best Regards,
Evgeniia

[Daniele Varsano]

Dear Evgeniia,
unfortunately the link you post just point to a blank excel file.
Anyway can you post the output files? I'm afraid that the damping parameter could affect the Haydock procedure, can you repeat also the same tests using a direct diagonalization for you simple tests and compare with the haydock procedure?

Best,
Daniele

[Daniele Varsano]

Dear Martin,

2) Another aspect which I like to share with you is that I used the sphere cut to avoid artificial screening, and noticed that IP energy oscillates with the number of G vectors!
I.e., when using a kinetic energy cut-off of 100 Ry and 140 Ry for the same super-cell size with sphere cut-off I have oscillations in my IP energy by almost 1 eV!
However, whenever I totally neglect the cut-off procedure in Yambo, my IP energy with respect to different kinetic energy cut-off energies becomes stable.

This is rather strange, the sphere cutoff potential has an analytic expression, and while it differs consistently at long range (small |G|) it tends to the usual 4pi/|G|^2 expression for large |G|. Can you post your results/reports ?

Thanks,
Daniele

[martinspenke]

Dear Daniele,

Sorry for my late response.
I actually deleted my data and gave up this topic completely, since the absolute energy positions were very dependent on the choice of
pseudo-potentials. I was unable to reproduce the PAW results from VASP with standard norm-conserving
pseudo-potentials for ionization potential and electron affinity of small molecular systems.

I hope in future YAMBO can handle PAWs, too.

Best wishes
Martin

[haseebphysics1]

Dear Daniele,

I was curious to know that can we set damping range something like this (negative second value):

% DmRngeXd
0.80000 | -0.20000 | eV # [Xd] Damping range
%

In fact, I want the large damping value at the start and it should be lower in the last part of the spectrum. So, does the above range make sense?


Regards,