parallelisation on various parameters

[manaswitakar]

Hi,

I have a few questions regarding convergence of certain parameters when using YAMBO. I am working on perovskites, and currently dealing with 2 different ones:
(i) cubic Cs2SnPbI6: this has 10 atoms in unit cell and has a distorted rock salt structure.
(ii) orthorhombic CsPbI3: this has 20 atoms in unit cell and has an orthorhombic Puma structure.
Both these structures are obtained from experimental sources.

I have followed the standard procedure for convergence at the DFT level (converging the total energy wrt energy cutoff and kpoints). For the cubic case (i), the converged ecut and kpoints are respectively 80 Rydberg, and 9x9x9 (resulting in 35 kpoints/qpoints). When I used these parameters for G0W0, I could do the calculation in YAMBO and got the almost desired results.

However, for the orthorhombic case (ii), I cannot go beyond 2x2x2 kpoints/qpoints (8 points); for 100 Rydberg; beyond this I am running out of memory. Also, in this case I am trying to use 1000 bands as the literature has done. In the literature, they did 1000 bands with 4x4x4 and 150 Rydberg cutoff.

My question is, how can the k-point parallelisation be done optimally? Is it also symmetry dependent (for cubic I could do more kpoints easily)? Also, I would like to know how to convert the RL units to Rydberg in the yambo input? For example, the literature uses 6 rydbergs for polarisability. What is the corresponding value in RL?

Please let me know as convenient.

Thanks and Best Regards,
Manaswita

[Daniele Varsano]

Dear Manasawita,

Also, in this case I am trying to use 1000 bands as the literature has done. In the literature, they did 1000 bands with 4x4x4 and 150 Rydberg cutoff.

please note that the number of bands strongly depends on the plasmon-pole approach used. Usually HL plasmon-pole needs more bands that GN ones, that it is the one used in Yambo. So it is possible you do not need such number of bands if your reference calculation has been done withe HL model.

My question is, how can the k-point parallelisation be done optimally?

In GW calculation, in particular for the Xo calculations, the memory is better distributed parallelizing on "c,v" bands rather than "k".

it also symmetry dependent (for cubic I could do more kpoints easily)?

Sure, the more symmetric is the system, lower will be the number of kpoints in the IBZ.

Also, I would like to know how to convert the RL units to Rydberg in the yambo input? For example, the literature uses 6 rydbergs for polarisability. What is the corresponding value in RL?

You do not need to convert it, Yambo will do. You can use Ry, mRy, Ha, mHa units in yout input file.

Best,
Daniele

[manaswitakar]

Thanks Daniele. I had the same impression that GN plasmon pole can be converged with lesser number of bands. However, the reference has done with much higher bands and they report quite a different result than what I get. Maybe these are other parameters as well the need to be checked. I cite the reference for your reference.
ACS Energy Lett. 2018, 3, 1787−1794.

Regards,
Manaswita

[manaswitakar]

Also, regarding the conversion to Ry from RL, I get an error message telling that initialisation is required in proper units. I am not very sure what it means.

Regards,
Manaswita

[manaswitakar]

Hi Daniele,

This is the error message I get when I try to use Rydberg.

[ERROR] Initialization not yet done. Use RL units in input or run initialization

Regards,
Manaswita

[Daniele Varsano]

Dear Manaswita,

as indicated by the error message, it is enough to rerun the inizialitaion.
> yambo -i
> yambo

eventually delete your ndb.gops in your SAVE directory before running the initialization.

About bands to be converged, it is a good practice to run your own convergence study.
In the paper you mention, it seems that calculations are done with Yambo so, you should no be problematic to reproduce them, please note that in the paper it is not just a G0W0 calculation: "Here we include eigenvalue self-consistency, within the simplified self-consistent scissor scheme".

Best,
Daniele

[manaswitakar]

Hi Daniele,

Thanks again for your response. The system under consideration by the paper (the same I try to reproduce) has 264 occupied states. The calculations are with spin-orbit coupling. I have tried to change the number of bands (at the nscf level) to see the convergence of the G0W0 band gap. However, this already converged at 300 bands and the difference between 2x2x2 and 4x4x4 k-grid was 60 meV. I checked also for 400 and 500 bands. The results did not change. So, I think it's safe to assume the gap converged at 2x2x2 and 300 bands.

Now, I have 2 questions:
1. Is 300 bands too low for consideration of unoccupied states if the system already has 264 occupied states (results converge though, maybe there is some other problem?)

2. If 300 or 400 bands with 2x2x2 is sufficient, then is it necessary to go to 1000? the calculation with 1000 and 4x4x4 is crashing even on 500 CPUs under my computational setup. Is there other parameter that I am missing somehow?

Please let me know at your convenience.

Best Regards,
Manaswita

[manaswitakar]

Hi Daniele,

I could achieve the results for CsPbi3, a given in the literature. I realised that the main ingredient in this issue was the NGBlks parameter, that needed careful convergence. Earlier I was not able to set this parameter to Ry units, and was using RL, the conversion was not working. Now, I can set the units in Ryd. I am still trying to find out the convergence with nbnd to check if lower number of bands do the job. In terms of k-points, I found I get exactly the same result with 2x2x2.

Thanks for your help.

Regards,
Manaswita

[Daniele Varsano]

Dear Manaswita,

the convergence wrt bands and kpoints are independent, so you can safely perform your convergence using the lower grid.
Bands and NGBlks instead are interdependent so you need to converge both simultaneously: 300 bands when 264 are occupied seems to me to be rather low, maybe you checked it using a low NGBlks value? You should repeat the convergence test with a higher value. If you cannot afford 1000 bands you can anyway to study intermediate number of bands and see the trend.

Best,
Daniele