NLCC on Yambo

[manaswitakar]

Hi,

I had a question on using NLCC containing pseudo potentials in Yambo. So, I have generated some pseudo potentials with semicore states in valence for Pb and I for doing calculations on CsPbI3. I used the ld1.x atomic code and I learnt that these pseudo potentials have non-linear core corrections in them. As a consequence, I use UseNlCC in the yambo input for the GW calculations. However, I read recently that NLCC should be avoided in YAMBO. So, can I trust my band gap values? I do get results similar to literature with this set up though.

Please let me know as convenient.

Regards,
Manaswita

[Daniele Varsano]

Dear Manasawita,

there is not a unique answer to your question:

1) If you use the NLCC reconstruction, you are doing the right thing to subtract the <Vxc> term from the KS eigenvalue, but then you do not know how to take them into account in the non-local part of the self-energy (e.g. HF part). This is the reason why also in QE there is a warning when using NLCC in combination with hybrid functionals.

2) If you neglect them, you know that you are doing an error in the <Vxc> reconstruction and you "hope" that this is compensated from their absence in the non-local term.

That's the reason why PP without NLCC (as for instance sg15) are preferred. You can anyway do a quick HF only calculation with and without NLCC inclusion and see how much they impact.

Best,
Daniele

[manaswitakar]

Hi,

Thanks for your reply. So, you mean I use the pseudo potential with NLCC and do a HF-only calculation in yambo with and without NLCC? There should not be any change in the DFT calculations right?

Regards,
Manaswita

[Daniele Varsano]

Dear Manasawita,

yes as you say. In this way, you can see the effect of NLCC in the <Vxc> term. Clearly, this will not solve the ambiguity on what's the better way to proceed.
You can then also compare your results with a sg15 psp which does not contain nlcc and does not suffer from the issue.
Best,
Daniele

[manaswitakar]

Hi Daniele,

Thanks again for your response. So, I did the tests as you suggest: I have 2 cases:

1. I have the experimental CsPbi3 structure that I use as benchmark. I do HF@LDA with SOC with the ld1.x pseudos (that have NLCC) with and without the 'UseNLCC' keyword. The results are as follows:

Without NLCC:

K-point Band Eo [eV] Ehf [eV] Vxc [eV] Vnlxc [eV]

1 296 0.000000 -1.834426 -13.08861 -14.92304
1 297 0.458391 2.666028 -11.13934 -8.931705

With NLCC:

K-point Band Eo [eV] Ehf [eV] Vxc [eV] Vnlxc [eV]

1 296 0.000000 -1.232656 -13.69038 -14.92304
1 297 0.458391 3.090614 -11.56393 -8.931705



2. With the same structure as above, I now do a G0W0@PBE (without SOC) using the ld1.x pseudos and the ONCV sg15 pseudos:

ld1.x: (here I use the UseNLCC keyword)

QP [eV] @ K [1] (iku): 0.000000 0.000000 0.000000
B=148 Eo= 0.00 E= -0.46 E-Eo= -0.46 Re(Z)=0.81 Im(Z)=-0.141559E-2 nlXC= -15.2426 lXC= -13.6432 So= 1.03643
B=149 Eo= 1.67 E= 2.08 E-Eo= 0.42 Re(Z)=0.82 Im(Z)=-0.137252E-2 nlXC= -7.25772 lXC= -10.1905 So= -2.42650

ONCV SG15:

QP [eV] @ K [1] (iku): 0.000000 0.000000 0.000000
B=148 Eo= 0.00 E= -1.39 E-Eo= -1.39 Re(Z)=0.81 Im(Z)=-0.137664E-2 nlXC= -17.1304 lXC= -14.1687 So= 1.25143
B=149 Eo= 1.63 E= 1.71 E-Eo= 0.08 Re(Z)=0.82 Im(Z)=-0.134488E-2 nlXC= -7.98805 lXC= -10.3729 So= -2.28870

The ONCV gives QP band gap of 3.1 eV compared to 2.5 eV with ld1.x. The all-electron limit G0W0@PBE band gap provided by Claudia Draxl is 2.2 eV. This way ld1.x wins over ONCV-sg15. This is my observation. Please let me know if I am missing anything.

Regards,
Manaswita

[Daniele Varsano]

Dear Manssawita,

from your data:

1) Test 1. HF gap using and not using nlcc is about 0.18eV, as said in previous post, both calculation suffer from some inconsistency, the difference I would say is not dramatic.

2) I would not have expected such large difference. Note, it is not guaranteed that calculations with different psp converge with the same parameters, in particular if the DFT energy cutoff needed to converge the ground state is not the same.

Be sure you have inserted all the G vectors (the one needed to converge the density i.e. 4*ecut_wfc in the evaluation of Vxc), I see that
<Vxc> in test 1 (NLCC) and test 2 same pseudo differ, I would investigate it.

Having said that, comparing with an all-electron calculation is a good benchmark.

Best,

Daniele