convergence of SCF calculations for metallic systems in CP2K

marci marc... at pci.uzh.ch
Tue Oct 4 10:05:57 CEST 2011


Dear Sanliang,

Yes, you can run simulations with metallic systems and yes normally
metallic systems are harder to converge.
Since k-point are not implemented with CP2K, meaningful results can
be
obtained only with large enough supercells.
For a metallic system, you have to consider the advantages of
smearing
the occupation number of the states around the Fermi energy.
We use a Fermi Dirac smearing where the temperature parameter sort
of
determine how many states are involved in the smearing, and can
affect
the convergence.
In order to do that the only possibility at the moment is to compute
at each step of the SCF minimization eigenstates and eigenvalues of
a
number of KS orbitals larger than the number of electrons, in order
to
include some unoccupied states as well. Haw many additional states
you
need depends on the system and on the smearing temperature.
This means that efficient algorithms as OT cannot be used, and you
have to resort to standard diagonalization, which scales much worse
with both system size and number of tasks.
Another useful technique is the mixing of the electronic density in
reciprocal space, to mend effects related to the charge sloshing.
Moreover, to get well converged states you might need to tighten the
convergence criterion (EPS_SCF).
All this can make the calculation quite lengthy.

A typical &SCF input section for a calculation with metals reads like:
&SCF
     SCF_GUESS  RESTART
     EPS_SCF 8.0E-7
     MAX_SCF 500
     ADDED_MOS  700
     CHOLESKY INVERSE
     &SMEAR  ON
         METHOD FERMI_DIRAC
         ELECTRONIC_TEMPERATURE [K] 300
     &END SMEAR
     &DIAGONALIZATION
          ALGORITHM STANDARD
     &END DIAGONALIZATION
     &MIXING
          METHOD BROYDEN_MIXING
          ALPHA   0.1
          BETA    1.5
          NBROYDEN  8
     &END
&END SCF

best
Marcella


On Oct 3, 5:28 pm, "S. Ling" <lsl.m... at googlemail.com> wrote:
> Hello all,
>
> I am preparing to start some new calculations for a very large system
> in which we have silver as substrate. I heard some comments saying
> that the SCF calculations converge very slowly for metallic systems in
> CP2K. I want to check if any of you have such experience with CP2K. If
> yes, could you tell me how did you deal with this problem? Any hints
> on special keywords to be set in the inputs? I also heard about some
> new developments of CP2K to address this problem. Could you give me
> some update suggestions?
>
> Thank you very much.
>
> Sanliang


More information about the CP2K-user mailing list