My Community

To which it may concerned,

For Optical properties, ALGO = Exact, I use the neither ICHG=11 or ISTART=1, reading the SCF WAVECAR or CHG*, it won't proceed. I delete the Ncore =4 or Change the NELMDL=1, or NELM=1. All of them do not works. Or it might be the fault of memory? I attach the related file in the file.rar. Thank you so much!

Optical properties
ICHG=11 or ISTART=1
ALGO = Exact
NBANDS = 496 (Set this parameters manually)
LOPTICS = .TRUE.
CSHIFT = 0.100
NELM=1
NEDOS = 2000
NCORE =4
LREAL= Auto
ISMEAR = 0
SIGMA = 0.01
EDIFF = 1.E-8
! LPEAD = .TRUE. (Write the derivative of the cell-periodic part of the orbitals)
NGX =224
NGY =224
NGZ =216


total amount of memory used by VASP MPI-rank0 1190325. kBytes
=======================================================================

base : 30000. kBytes
nonlr-proj: 171357. kBytes
fftplans : 167183. kBytes
grid : 759141. kBytes
one-center: 2052. kBytes
wavefun : 60592. kBytes

INWAV: cpu time 0.0000: real time 0.0001
Broyden mixing: mesh for mixing (old mesh)
NGX = 75 NGY = 75 NGZ = 69
(NGX =448 NGY =448 NGZ =432)
gives a total of 388125 points

initial charge density was supplied:
charge density of overlapping atoms calculated
number of electron 814.0000000 magnetization
keeping initial charge density in first step


--------------------------------------------------------------------------------------------------------


Maximum index for non-local projection operator 20050
Maximum index for augmentation-charges 2222 (set IRDMAX)

Hi,

496 bands (NBANDS=496) is a lot for the exact diagonalization algorithm (ALGO=Exact). The system is quite large (132 ions) and there is only 1 space-group-symmetry operation. At least you only have 1 k point.

Please note the following warning was written to the OUTCAR: So, the number of bands in your calculation is actually 504. This is a consequence of running on 24 cores. 504 / 24 = 21

Can you please try a smaller system to check that your setup works in principle?
How long did you wait until aborting the calculation? You might simply need to run on more cores, if that is available to you. The parallelization of VASP over bands will speed up your calculation significantly.

Best regards,
Marie-Therese

Ps: I noticed that LMAXMIX is not set to 4, although your structure contains d orbitals. Please have a look at the description of LMAXMIX on the Wiki to find more information.

Thank you for your quick reply! I had did the example of Si calculation and setup can finish the job successfully.
Regarding to this system, for one thing, I did one scf calculation before and grab the NBANDS from it. To my surprise, NBANDS changed.
For its memory and other thing, I had think of it but did not consider it as the main factor. Should I reduce the bands amount so as to reduce the calculation and how to reduce it? (The instruction of vapswiki just told me to keep the same NBANDS. Or Other instruction told me to double or triple the bands.) I found some suggestions from NELM=1 might reduce the calculation amount. However, my system is still very large through I only choose one K point. Maybe the system is too large, I found the OSZICA just stay at certain point and OUTCAR cannot change for 10 more minute. So I kill the calculation and try another parameter. I will try to give them more cores and wait more time and check whether it works or not.

Good, so the Si example works.

In the SCF calculation, how did you submit the job? Probably not on 24 cores, but on 16 cores?

This is bad news, but LOPTICS includes a sum over unoccupied bands! Consequently, the number of bands (NBANDS) should be increased significantly with respect to the default value. (The default of NBANDS aims to include a minimum number of bands while including all occupied bands. That is because the DFT total energy only depends on occupied bands and is minimized to obtain the ground state Kohn-Sham orbitals.)

To ensure you have included sufficient bands, it is expedient to perform a convergence study:
- Increase the number of bands (say NBANDS=512), also better set LMAXMIX, run on sufficient number of cores (perhaps 64 cores) starting from the ground state WAVECAR (with 496 bands)
- Next, increase the number of bands even more (say NBANDS=576 on 72 cores) and start from the WAVECAR with NBANDS=512.
- Check if the desired quantity is converged: Is the frequency dependent dielectric function of both calculations the same? If not, increase the number of bands further.

It also sometimes helps to have a look at the band structure and think about the electronic configiuration to estimate the number of bands that will lie just above the Fermi energy. For instance, if there is 1 electron occupying a p orbital, the 5 unoccupied p orbitals are probably just above the Fermi level. As your system has many fully occupied d orbitals I assume that the number of bands does not have to be increased more than I suggested above, but it needs testing.

Anyways, it can be confusing how to set NBANDS for optical response calculations, because there are multiple algorithms implemented for the static and frequency dependent dielectric function. If you point out which part on the VASP Wiki is confusing, we will try to improve the description.

Best regards,
Marie-Therese

Hi, Marie-Therese,

Thank you for your quick reply!
For SCF calculation, I used on the same cores as Optics calculation. Maybe I should request more cores to finish the task.
I will try to proceed the following advices including the NBANDS and LMAXMIX as you recommend and check the further result.

Thank you very much!