Dielectric with hybrid functional

[IBRAHIM]

Dear All,
I am calculating frequency dependent dielectric properties using hybrid functional and SOC for doped-perovskite conventional cells.

1- For exact diagonalization to get virtual empty states of sufficient quality, I tried ALGO=Exact but it is not working due to memory issues (forrtl: severe (41): insufficient virtual memory), even with ulimit -s unlimited

So, I continue with ALGO=All (with LREAL =.FALSE. due to the HF part with optical routine) and WAVECAR of normal run (with LREAL=.TRUE.) is it OK without losing accuracy? I have calculated a system with 5-7 atoms and ALGO Exact works for that system! 

2-For SOC, I first double number of bands (to be 384), then for optical, I doubled again (768) but I got this error:
"reading WAVECAR reading wavefunctions of collinear run, up
 ERROR: while reading WAVECAR, plane wave coefficients changed       23031        23031"

It seems due to the very high number of bands, so I decreased the number to 576 bands and it is finished for pure system, but strangely the doped systems fail while calculating the dielectric response at this step:
"writing wavefunctions
 optical routines
 imaginary and real dielectric function
 recalculating local potential from charge density
 direction            1
bash: line 1:  7746 Killed "

I repeated it two times to check if it is due to the server but it failed with the same message. Also, I doubled the CPU (192 cores) without success. So I think it is due to difficulty in the computation of the derivative of the cell-periodic part with perturbation theory. But strangely it is finished for the pure system (even it took more than 2 days to calculate the dielectric function after writing wavefunctions part) and all settings are the same in all cases.

3-So, with the previous setting of step 2, I tried the finite method LPEAD=.TRUE. (also with NBANDS=576) because I have non-metallic systems, and it is finished for both pure and doped systems. So please what happened in that case, Does LPEAD run only on occupied bands? If so, is a high number of empty states still required for LPEAD?


Many thanks in advance,
Ibrahim

[IBRAHIM]

Frequency dependent dielectric calculations at the independent-particle approximation level.

[merzuk.kaltak]

Dear Ibrahim,

Is there a reason you set LSCALAPACK=.FALSE. in the INCAR?
For your system you certainly want to run VASP in parallel with MPI and scaLAPACK support.
This is probably the reason why you run into memory issues.

[IBRAHIM]

Thank you for your reply.

I use it as save option for the server.

However, I have tested a normal DFT optical job without it and the same issue.

One more thing, for systems 5-7 atoms, even with the same system but with perimitive cell of 10 systems I did not run to these issues and I used ALGO Exact and perturbation theory for the derivative of the cell-periodic part, ofcourse takes long time, but not failed directly as happened with my current jobs.

PS. For step 1 in my first post, I mean LREAL=A (not true).


Best regrads,
Ibrahim

[merzuk.kaltak]

Dear Ibrahim,

scaLAPACK is required for large cells (like the one you are studying).
Please submit the makefile.include that was used to compile VASP. I suspect your vasp installation is not supporting scaLAPACK.

[IBRAHIM]

Thank you for the recommendation.

OK it is attached.

Best regards,
Ibrahim

[merzuk.kaltak]

The makefile seems OK.
Which vasp-version do you use?
Also, can you attach the POTCAR?

[IBRAHIM]

Thank you.

I use vasp.5.4.4 and also tested vasp.5.4.1.

Attached is POTCAR.

Best regards,
Ibrahim

[merzuk.kaltak]

Hi, I realized I do not have the doped structure.
Can you send me the POSCAR and POTCAR for the doped system please?

[IBRAHIM]

Dear Dr. Merzuk.kaltak,

Thank you very much. Please find the attachment.

In between, I have tested scaLAPACK but unfortunatly the same error for algo exact (memory issue from the begining) and for algo all (for dielectric function calculation by perturbation theory after scf steps in doped system).

So, I followed suggestions here forum/viewtopic.php?f=4&t=18475

but scaLAPACK works for small cell and also the calculations work for the conventional cell but with single kpoint with NBANDS 576 (but same error for 768).

So, it seems the large kpoint number could affect. I recompiled VASP again and still checking the calculations once more for conventional cell and my KPOINTS file.

Moreover, my error with very large number of NBANDS is reported here forum/viewtopic.php?f=4&t=18444
but the disscussion did not complete.

In fact I am interesting in dielectric function at independent particle approximation level at energy range near the bandgap,
so I follow suggestions forum/viewtopic.php?f=4&t=18131
but with algo All (instead of ALGO normal in that link) and LREAL=.FALSE. (due to HF part) in addition to LPEAD=.TRUE. to avoid the last error in doped system due to the derivative of the cell-periodic with perturbation theory .

Best regards,
Ibrahim

[IBRAHIM]

Dear All,

Update:

I recompiled VASP and still the above issues are found with the new installed version.

However, I tried ALGO=Normal as mentioned in the forum https://www.vasp.at/forum/viewtopic.php?f=4&t=18131

and the optical routine with perturbation theory and NBANDS 576 is finished without errors, here the steps:

Step1: Normal HSE: ALGO=All, LREAL=A, NBANDS=384
Step2:  
LOPTICS=.TRUE. 
ALGO=Normal 
LREAL=.FALSE. 
NBANDS=576
CONTCAR and WAVECAR from step1 
NELMIN =5 (I enforced vasp to do minimum steps to allow more convergence of new bands) 

In step 2 I switched to LREAL=.FALSE. because I got a message that: LOPTICS is not supported for HF part and LREAL =A , so switch to LREAL=.FALSE.

Best regards,
Ibrahim