HSE06 calculations in VASP 6.6.0 vs VASP 6.3.0

[shivam_sharma]

Hi,

I used vasp 6.6.0 recently with HSE06 hybrid exchange functional. The total energy calculated by vasp 6.6.0 is off by approximately 30eV in comparison to vasp 6.3.0. This affects the band gap and forces significantly. I saw that all other energies are same but exchange energy is off by 30eV and that is reflected in the total energy of the system as well.

I have used same settings for both the versions. What could be the source of this offset?
Let me know if you need more information.

Many thanks,
Shivam

[michael_wolloch]

Dear Shivam Sharma,

Thanks for reaching out.

Please provide all your input and relevant output files (at least OUTCAR) for both your calculations, as per the posting guidelines.

In this case, it will also help if you provide the toolchains you used for your builds, and the makefile.include files if you have them.
Without this information, it is impossible to know what is going on.

We do have a sizable number of regression tests in our validation tests that check hybrid functionals, including HSE06, so you can also run those tests if you want. Set:

Code: Select all

export VASP_TESTSUITE_RUN_HYB=Y

in your .conf file in the testsuite folder to select only the hybrid functional tests.

We, of course, validated VASP 6.6.0 against the full test suite before we released it.

Cheers, Michael

[shivam_sharma]

Dear Michael Wolloch,

Thanks for your response.

I have attached the compressed directory containing input and output files of both the versions. The suffix to the file name tells the version of vasp used.
For example INCAR.vasp6p6p0 is input file for vasp 6.6.0.
I have also attached the output file from sbatch for each case.

I don't have the makefile.include as I am running it on HPC. I have attached the submit files for both the versions, if that helps.

Please let me know if you need any further information.

Many thanks,
Shivam

[michael_wolloch]

Dear Shivam,

Looking at your INCAR and OUTCAR files, you do not use the same settings.
You have different settings for both ALGO and LKPOINTS_OPT.

Furthermore, your OUTCAR for 6.6.0 does not match your INCAR for 6.6.0:
INCAR.vasp6p6p0:

Code: Select all

#ENCUT = 520
PREC   = Accurate
EDIFF  = 1E-6
NELM   = 50
KPAR   = 40
IBRION = -1
NSW    = 0
ISIF   = 0
ISTART = 1
LORBIT = 11
LMAXMIX = 4
MAGMOM = 18*0.0
LKPOINTS_OPT = .TRUE.
ISYM = -1
LHFCALC = .TRUE.
GGA     = PE
AEXX    = 0.25
HFSCREEN = 0.2
ALGO    = Normal
TIME    = 0.4
PRECFOCK = NORMAL
LSORBIT = .TRUE.
LNONCOLLINEAR = .TRUE.
HFRCUT = -1
ISMEAR = 0
SIGMA  = 0.050000
NBANDS = 80
LREAL  = .FALSE.
LASPH  = .TRUE.
ADDGRID = .TRUE.
LWAVE  = .TRUE.
LCHARG = .TRUE.

OUTCAR.vasp6p6p0:

Code: Select all

 INCAR:
   SYSTEM = WTe2 monolayer x strain +1.00% HSE band
   PREC = Accurate
   EDIFF = 1E-6
   NELM = 50
   KPAR = 40
   IBRION = -1
   NSW = 0
   ISIF = 0
   ISTART = 1
   LORBIT = 11
   LMAXMIX = 4
   MAGMOM = 18*0.0
   LKPOINTS_OPT = .TRUE.
   ISYM = -1
   LHFCALC = .TRUE.
   GGA = PE
   AEXX = 0.25
   HFSCREEN = 0.2
   ALGO = Normal
   TIME = 0.4
   PRECFOCK = NORMAL
   AMGGAX = 0.0
   AMGGAC = 0.0
   FOCKCORR = 2
   LSORBIT = .TRUE.
   LNONCOLLINEAR = .TRUE.
   HFRCUT = -1
   ISMEAR = 0
   SIGMA = 0.050000
   NBANDS = 80
   LREAL = .FALSE.
   LASPH = .TRUE.
   ADDGRID = .TRUE.
   LWAVE = .TRUE.
   LCHARG = .TRUE.

Note that AMGGAX, AMGGAC, and FOCKCORR are set here. They are not set in the INCAR, and neither in the OUTCAR.vasp6p3p0, instead that OUTCAR shows Wannier90 tags that are not in the INCAR.vasp6p3p0. Please review all your files. I strongly suspect that you will not find different energies when you don't do different calculations.

I used the 6.3.0 inputs (reduced the kpoint density by 50% to have only 32 kpoints in total and make the computation less demanding) and ran on both 6.6.0 and 6.3.0. There is no difference in energies (see attached OUTCARs).

Please use the same inputs for both calculations and report back if you still observe different energies.

Cheers, Michael

[shivam_sharma]

Dear Michael,

Thanks for looking into it. I understand the difference in settings. I was trying different parameters to match the two results.

In the attached files, I used exactly the same parameters as in your OUTCAR files. The final energy for the vasp 6.6.0 is extremely close your answers. However vasp 6.3.0 final energy is still very different. The difference is coming because of exchange energy component.
I stopped output from vasp6.3.0 before the final diagonalization after seeing that the energy has converged to different energy. You can see that in myjob.o18928973 file.

I also saw that you used 6.3.2 version for testing. Was there a bug in 6.3.0 or is it because of different built environment for two versions of vasp in HPC? I have attached the bash file as well, just in case.

Sincerely,
Shivam

[michael_wolloch]

Dear Shivam,

Thanks for uploading the correct files. This might be known issue 80, which was fixed in 6.3.1. I used release tag 6.3, which gave me version 6.3.2, sorry for not checking.

Please re-test without setting HFRCUT. If that solves the issue, we have found the bug.

Cheers, Michael

[shivam_sharma]

Hi Michael,

I ran it again without HFRCUT and answers are matching. Great!

If we switch off HFRCUT, will the band diagrams from vasp6.3.0 be trusted? As mentioned https://vasp.at/wiki/Band-structure_cal ... unctionals that HFRCUT = -1 converges best for systems with a band gap. If HFRCUT
is set other than -1 it will cause discontinuities in band-structure calculations.

Thanks,
Shivam

[michael_wolloch]

Hi Shivam,

Great news. Now we know what went wrong. I think the obvious solution here is to use VASP 6.6.0, since you have access to it. If it is not available on one of your clusters, you can at least perform a comparison with and without setting HFRCUT with 6.6.0 and then decide if your results without it are accurate enough to use 6.3.0. Or you can compile 6.6.0 yourself.

I think this issue is solved, but if you have further questions, please don't hesitate to ask. If not, please let me know as well, so I can lock the topic.

Cheers, Michael

[shivam_sharma]

Hi Michael,

Thanks once again for looking it. I have no further questions. You may lock the topic.

Best,
Shivam