Box cutoff

[wachr]

Dear yambo developers,

I try to use the box cutoff for a given system and it does not yet become clear to me, whether the box cutoff parameters in yambo 3.4.2 are still half the box size (like in 3.4.1) or the entire box size. For my system, the question is whether I have to take

%CUTBox
9.1962675|9.1962675|0.0| # [CUT] [au]
%

or

%CUTBox
18.3925350|18.3925350|0.0| # [CUT] [au]
%

Kind regards and thanks for your answer!
Christian

[Daniele Varsano]

Dear Christian,
the parameter are exactly the same of 3.4.2. As a rule of thumb, just a little bit smaller than your cell.
In this period we are testing other ways to eliminate spurious interactions between replicas and hopefully the input
part governing the coulomb cutoff will be soon quite a lot simplified.

Best,
Daniele

[wachr]

Dear Daniele,

thank you very much for your helpful feedback! For box cutoff, we use about half the cell size, so I guess we are save.

I wish you and the team success for implementing an improved Coulomb cutoff! This topic in general gave us headaches often of a times . (See the P.S. below). However, the support here is really great.

Kind regards,
Christian

P.S. We face one issue that I want to adress in a further, more detailed post (with input and output files) - after our box-cutoff calculations will be finished:

In a periodic 1D system, cylindrical cutoff seems not lead to convergence using 1x1xN k-grids - neither for GW nor for BSE (as RIM becomes ineffective, then). With RIM and without cutoff, we achieve convergence - however, no supercell convergence (as is known). Finally, we found out that using 2x2xN grids, convergence can be achieved with cylindrical cutoff - without the use of RIM. Calculations are just 3 times more expensive, then.

So I am curious about the results on the box cutoff (+ RIM + 1x1xN k-sampling) due to its different implementation. (which was somehow ignored in GW/BSE in yambo 3.4.1rev51. But in 3.4.2, it seems to work.)

[Daniele Varsano]

Dear Christian,

For box cutoff, we use about half the cell size, so I guess we are save.

Beware on that, it depends how big is your cell with respect the size of your system. Half of the cell could be very small, as it means an interaction ranging up to Lcell/4 so you cut interaction inside the cell. The Lbox size=cell_size finally it means half cell and it is safe. So please check this as it could be crucial.

About your comment, I know that the coulomb cutoff issue is alway painful.
Anyway in order to clarify:
Cylinder shape do not need RIM, you can include it but it is ignored as you have semi-analytical expression and integration over BZ fo build up the potential is not needed.
Box shape: RIM is mandatory as it is needed to build up the potential.

About the cylinder on 1D system, it should be quite effective. Somehow it needs more k point along the axis, it depends on the system, but it sounds be a bit strange you need to sample in the other directions, unless you have small vacuum.

Anyway, I'm currently testing a new strategy that should simplify all this stuff for both 1S and 2D systems, at least for orthorombic cells. Up to know the tests are encouraging.

Best,
Daniele

[wachr]

Dear Daniele,

Beware on that, it depends how big is your cell with respect the size of your system. Half of the cell could be very small, as it means an interaction ranging up to Lcell/4 so you cut interaction inside the cell. The Lbox size=cell_size finally it means half cell and it is safe. So please check this as it could be crucial.

Hmm, I guess I ran into this pitfall - even though I should have been aware of that. Thanks a lot for particularly pointing that out!

About the cylinder on 1D system, it should be quite effective. Somehow it needs more k point along the axis, it depends on the system, but it sounds be a bit strange you need to sample in the other directions, unless you have small vacuum.

I know it should converge wrt. to the k-points. I'll send the calculations soon and the results, such that any issue is visible.

Anyway, I'm currently testing a new strategy that should simplify all this stuff for both 1D and 2D systems, at least for orthorombic cells. Up to know the tests are encouraging.

I am looking forward to trying this out, one day!

I think it would be great for the user to have an output - right at the start of the calculation - (to be read by gnuplot or other plotting tools) that simply prints the supercell, the border of the cutoff potential and the maximum allowed cutoff potential for the supercell (centered around the cell center) to be sure that the right calculation is on its way. Just a suggestion - I know that you are working on yambo during your spare time and I highly appreciate that!

Kind regards,
Christian

[Daniele Varsano]

Dear christian
you can generate an output with cutoff potential in real and reciprocal space . It is a bit rough but can be useful.
yoy need to uncomment a variable named CutColtest ir similar. Now I am not in front of the laptop and I cannot check the exact name. It appears adding -V RL when building the input. If not add -V all.
Best
Daniele

[wachr]

Dear Daniele,

thanks a lot for this hint! It worked and I got some good estimate on the cutoff potentials.

About the cylinder on 1D system, it should be quite effective. Somehow it needs more k point along the axis, it depends on the system, but it sounds be a bit strange you need to sample in the other directions, unless you have small vacuum.

I see now that the convergence of the gap wrt. the number of G vectors is crucial. The number of Gvecs seems to depend on other parameters such as number of bands, kpoints, w/ or w/o truncation. In the non-truncated case, I converged the GW gap for 20'000 Gvecs for a given 1x1xN k-point set. However, with other parameters such as a different k-point sampling in the non-truncated case, the gap started to scatter in a certain range, which is a hint of underconvergence - but I never knew exactly, what.

Now, I am using 80'000 Gvecs in the exchange sum and the calculations seem to converge. Thus, the Gvecs are crucial for Coulomb-truncated calculations (!).There was the picture in my mind that a few Gvecs are often enough (once converged) and that GW results are rather insensitive to this number - therefore, I never thought about this possibility. But now, it seems that this is the key for the convergence of the GW gap.

I will carefully check this statement for other according BSE calculations (see a much older post of mine). In that case, we later found a solution by the use of 2x2xN-grids in the combination with Coulomb truncation for the BSE. But this might be just a workaround rather than a real solution.

How about a "Running yambo -> Convergence (issues)" section in the Forum? [Such as the great posting by Claudio Attacelite. Should be a "must read" fro all users!]

Kind regards,
Christian

[Daniele Varsano]

Dear Christian,

the number of the needed Gvectors is strongly system dependent as it reflects the non-homogeneity of the system.
Usually for the Exchange part of the self energy a lot of G vectors are needed, but this is not a very time consuming part of the calculation.

When doing calculations w/o cutoff with non 3D k-point sampling using RIM is mandatory otherwise you will never converge wrt k points, actually your
Sigma_x will diverge when adding k points.

A subforum "Convergence issue" is a good option, I will open as soon I find the time.

Many thanks,

Daniele

[wachr]

When doing calculations w/o cutoff with non 3D k-point sampling using RIM is mandatory otherwise you will never converge wrt k points, actually your
Sigma_x will diverge when adding k points.

Right. Using RIM, the spectra converge by the use of 20'000 Gvecs . This was not made clear in my post.

A subforum "Convergence issue" is a good option, I will open as soon I find the time.

Great! I volunteer on writing about some issues in low-dimensional systems, which took us some time to understand.

[Daniele Varsano]

Dear Christian,
I've created a "convergence issues" subforum.

Thanks for your time!


Daniele