instability on scf during geoopt

Jun chen... at googlemail.com
Thu Mar 25 10:11:35 UTC 2010


Thanks, Matt.

I noticed that as well. I was trying to understand the problem I
described above.
In fact, the situation is more complicated. The results are dependent
of the scheme to optimize wavefunction.
When using matrix diagonalization, the results of restricted KS and
low spin (multiplicity 0) unrestricted KS are the same. The lowest
energy is actually given by high spin unrestricted KS (i.e.
ferromagnetic). In contrast, the most stable state using OT is low
spin unrestricted KS, giving antiferromagnetic ordering. This is in
line with the result from my colleague who used another code which is
able to specify spin states of individual atoms. Although the system
is quasi-metallic (metallic or with small bandgap, depending on RKS or
UKS, and geometry), OT seems work for some reason I don't quite
understand. Possibly, it helps that there are very few near degenerate
states close to fermi level.

Cheers,
Jun

On Mar 24, 4:50 pm, Matt W <mattwa... at gmail.com> wrote:
> Hi Jun,
>
> I do see the behaviour you describe.  What I tried to hint at above is
> that you'll converge (to the same solution) if you use a restricted
> wavefunction.  I've seen this for another system of my own.
>
> I'm not sure if it is just the smaller number of degrees of freedom
> that help the non UKS converge, whether the settings for the mixing
> should be altered for UKS or whether there may be something else.
> I'll try and cut down my system too.
>
> Cheers,
>
> Matt
>
> On Mar 19, 6:40 pm, Jun <chen... at googlemail.com> wrote:
>
> > Thanks for the reply, Matt.
> > I am afraid that I cannot get the first scf converged for a smaller
> > cell. Don't know if it is related to the lack of k-point sampling
> > being inappropriate for a small system.
> > I keep trying to find a smaller system having the same problem.
> > Anyway, the present system is not that big, I am grateful if anyone
> > being able to reproduce the problem can help.
>
> > Cheers,
> > Jun
>
> > On Mar 19, 5:15 pm, Matt W <mattwa... at gmail.com> wrote:
>
> > > Can the problem be made smaller?  Is it enough just to use UKS and a
> > > small system to get convergence problems?
>
> > > Matt
>
> > > On Mar 19, 2:19 pm, Jun <chen... at googlemail.com> wrote:
>
> > > > Hi all,
>
> > > > I have been testing on some metallic system using diagonalization
> > > > method. I used a set up found in a previous post with small changes.
> > > > It works well for the first scf run; wavefunction got converged with a
> > > > few tenth steps. But in the following geoopt, wavefunction cannot get
> > > > converged. I played a bit with extrapolation method (PS and ASPC) and
> > > > order, and it didn't help. Any suggestion and help are much
> > > > appreciated. My input file (td.inp) is uploaded.
>
> > > > Thanks.
> > > > Jun



More information about the CP2K-user mailing list