GW in defect states

[sitangshu]

Dear Yambores,

I came across a topic on defects in a bulk system. Can a GW calculation using yambo be done for that? If I take a supercell approach in DFT, say 5 x 5 size, how should I proceed for GW convergence tests?

Thanks and regards,
Sitangshu

[claudio]

Dear Sitangshu

GW can be used to study defects, but converge is very difficult.
See for instance:
https://journals.aps.org/prb/abstract/1 ... .83.144115

You have to perform calculations with different supercells and check that defect level positions do not change anymore in GW.

Moreover, there is another problem, some defects are open-shell systems, and therefore GW results are intrinsically wrong,
for a discussion see here:

https://journals.aps.org/prl/abstract/1 ... 109.036406
https://www.nature.com/articles/s41535-018-0103-6

with best regards
Claudio Attaccalite

[sitangshu]

Dear Professor Claudio,

Thank you for your quick answer. I can imagine the convergence issue ! Let me try with GPU version of yambo. I look forward to learn yambo using gpu in the upcoming webinar, and will get back to this discussion.

Thanks and regards,
Sitangshu

[Daniele Varsano]

Dear Sitangshu,

if you want to study doped systems, in this case, supercell size reflects different doping concentration.
You can have a look at this paper:
https://journals.aps.org/prmaterials/ab ... s.3.045401

For a correct description of the defect states anyway, difficulties can arise from the DFT approximation used as a starting point, both structural (polarons) and electronic (localization) and this should be investigated with care.

Best,
Daniele

[sitangshu]

Many thanks Daniele for your suggestion.
I realized the main convergence criteria are supercell geometry and huge number of unoccupied states and for BSE number of transition bands. Right now, the structure is not open-shell system. So, I hope GW can be performed on this. Before even a trial run, can you hint me :

1. I have 2 GPUs Tesla V 100
2. I have a 8 node 320 core HPC.

Which one would be more viable to work with? Expected number of atoms for say a 5 x 5 geometry should be around 150.
Is this architecture worth for a 10 x 10 supercell? Any time constraint, for eg: number of core hours?

Regards,
Sitangshu

[Daniele Varsano]

Dear Sitangshu,
150 atoms is a rather big system. According to my experience, Yambo is very fast on GPU, but you can encounter a problem in memory allocations which is bound by the memory of the GPU (16GB I presume), so you would need to distribute memories and you need more cards. Of course, this dependent critically by the k points sampling you need to adopt. Briefly, this is just a warning as I do not have a clear answer to your question, maybe someone else can give you advice on this respect.

Best,
Daniele

[sitangshu]

Thanks for your answer, but I foresee a 5 5 1 supercell with 2 2 1 k-points should be fair enough?

Thanks?

[andrea.ferretti]

Dear Sitangshu,

thanks for raising this point (how to choose among CPU and GPU-based computational resources for what concerns yambo calculations).
I think this is a very important one.

Daniele has already mentioned the key point: GPUs are very fast but have little memory, and this is critical for a typical yambo run.
Considering a 150 atom system, given the experience we had so far if I had to bet I think 2 V100 cards would not be enough, and therefore I would go
for the CPU options.

In particular, when running on GPUs I think it is very important to have a flexible environment such that you can increase the number of GPUs you
have access to. This is exactly to fit your calculation into the available memory. Once this is done, the calculations are extremely fast.

Andrea

[sitangshu]

Many thanks Professor Ferretti for your valuable comments. I am trying a test.... I will comment on the developments...

Regards,
Sitangshu