Co(111) Geometry optimization with kpoints

[Manuel Corral]

Dear Colleagues,

I have been trying to run a geometry optimization for with a Co(111) 4-layer slab model without success. 

I set up my initial input file according my previous experience with another software (VASP) and then changed different calculation settings: number of additional MOS (10, 30, 100), electronic temperature of the smearing (300 K, 500 K), decreasing the cell in the vacuum direction, changing the number of kpoints (gamma point with an increased cell), the Poisson solver (analytical vs non analytical), the diagonalization algorithm, settings in the mixing algorithm ... nothing worked. The SCF cycle never converges.

I would very much appreciate if you could give some advice or feedback on my input files (enclosed).

Thank you in advance for your help.

Kind regards,

Manuel Corral

[fabia...@gmail.com]

Dear Manuel,

I changed a couple of things in the input and it's converging now (attached below). The changes include:

To use the WAVELET poisson solver I rotated the slab such that B is the surface vector.

I also decreased RELAX_MULTIPLICITY because large values hinder scf onvergence. Maybe you need to test various initial guesses (see here: https://manual.cp2k.org/cp2k-8_1-branch/CP2K_INPUT/FORCE_EVAL/SUBSYS/KIND/BS.html) to find the minimal energy configuration.

I increased the ELECTRONIC_TEMPERATURE in the &SEMEAR section. 300 K corresponds to just 0.02 eV in VASP which is very little.

Cheers,

Fabian

[Manuel Corral]

Dear Fabian,

Thank you very much for the commented file. It works, which it is already a huge improvement, though it may take about 75 electronic iterations for each ionic step, which may render a MD calculation too expensive. I wonder if there is anything that could be done to speed it up.

By the way, why do  you need to rotate the cell for the Wavelet Poisson solver ? Based on this rotation, I would have thought of changing the Monkhorst Pack set of k-points from 7x7x1 to 7x1x7, which you didn't and, when I tried it, it did not work properly:

 *** WARNING in pw/ps_wavelet_methods.F:236 :: Density non-zero on the ***

 *** edges of the unit cell: wrong results in WAVELET solver.

Thanks again.

Regards,

Manuel

[fabia...@gmail.com]

Hi Manuel,

According to the manuel the wavelet solver supports "slab-boundary conditions (but only PERIODIC XZ)." You are correct to change the k-point grid accordingly.

To avoid the warning you mentioned need to increase the cell length in the y-direction.

To reduce the number of scf iterations you can optimize the &MIXING section, in particular NBUFFER and ALPHA. If you don't expect the multiplicity to change during MD you can disable RELAX_MULTIPLICITY to further improve scf convergence. And lastly you can tune the EXTRAPOLATION.

Cheers,

Fabian

[Marcella Iannuzzi]

Dear Manuel, 

You can avoid the Wavelet Poisson solver and use the surface dipole correction instead

     SURFACE_DIPOLE_CORRECTION

      SURF_DIP_DIR Z

When the optimisation method is the diagonalisation, the ASPC is not a good extrapolation method.

For metals and G-space mixing, the best choice is probably USE_PREV_WF

Running MD for such a small cell using k-points sounds a little bit weird to me. 

Kind regards

Marcella

[Manuel Corral]

Dear Fabien and Marcella

Thanks to both of you for your replies. 

Yes, my slab is small indeed. At this very moment I am just learning to use the code for this kind of systems and benchmarking results against to what I am familiar to. In a production run I will use a significantly bigger cell. But I was planning to do that next, after I manage to fine tune the SCF convergence with smaller/faster systems.

Best regards,

Manuel