Is it possible to study/converge magnetic surfaces? repost..

marci marc... at pci.uzh.ch
Fri Oct 28 12:20:58 CEST 2011


Dear  Valerio

I did a second test with a 6x6 slab and 6 layers, and by initializing
the multiplicity to 131,
I get a reasonable distribution of spins. Atoms belonging to the same
layer have all the same spin moment, the outermost layer have the
largest spin moment 0.66, and the two innermost layer have the lowest
spin moment 0.56.
It takes quite a number of iterations (about 140) to converge with a
convergence criterion of 1E-7.
I used a mixing parameter of 0.08 and an electronic temperature of
2000K.
The RELAX_MULTIPLICITY keyword is not needed when the smearing is
used.
The occupation of the states is attributed according to the Fermi-
Dirac distribution, through the evaluation of the Fermi energy at each
SCF step.
This means that the fractional occupation numbers can change, as well
as the number of electrons per spin channel, only the total number of
electrons is constant.
In the specific case, I monitored the number of electrons per spin
channel, and it remains quite stable from the beginning to the end,
though there are some fluctuations, in particular during the first
iterations.
The final difference between spin up and spin down is of 131.8
electrons.

best
Marcella

On Oct 26, 4:08 pm, Valerio Bellini <valerio... at unimore.it>
wrote:
> Il 26/10/11 15.30, marci ha scritto:
>
>
>
>
>
>
>
>
>
> > Dear  Valerio,
>
> > I tried your system, 5x5 Ni(111) slab, and I could converge the
> > electronic structure by using more or less the same settings that were
> > in your input.
> > It needs many iterations and the energy keeps oscillating for a long
> > time before the algorithm can find a good minimum.
> > However, what is really annoying, is that at the end the electrons are
> > redistributed between the two spins in such a way that the final
> > magnetization is zero, in spite of the fact that the initial guess had
> > a high multiplicity.
> > It seems that with the present settings and system size, the algorithm
> > finds a minimum with no magnetization, and this should be also the
> > reason why starting from a magnetization different from zero it takes
> > such a long time to converge.
> > It is possible that one problem is the size of the system. One should
> > check larger boxes to verify that.
> > What I can tell for sure is that the optimization of the bulk (216
> > atoms) electronic structure, gives the expected magnetization (~0.6
> > magneton per unit cell) by using more or less the same SCF set up.
>
> > best
> > marcella
>
> Dear Marcella,
> Thank you for the answer.
> Two comments:
> 1) I did calculation for the same system, using Gamma point only, with
> another code (VASP),
> and the total magnetic moment of the cell relaxed to around 86 bohr
> magneton.
> With a better multiplicity guess I thought convergence might be easier,
> but it was
> not the case.
> If I try 87 as multiplicity and I run more than 500 iterations with
> Broyden,
> the system converges using Diagonalization+Broyden up to 0.005 Hartree,
> but as said in the previous e-mail, the magnetic moments are not equal
> for different atoms
> in the same plane, so the system in reality is far from convergence.
> Could I ask you how many iterations did it take to you?
> 2) I do not understand how you could get a non-magnetic solution.
> If you impose the multiplicity to some value, and you don't allow
> relaxation of it
> (using the keyword, RELAX_MULTIPLICITY) the multiplicity of the system
> should remain
> constant (like in a fixed spin moment calculation).
> So this means that you inserted that flag in the input file, is that
> correct?
> thanks,
> Valerio


More information about the CP2K-user mailing list