[gpaw-users] Converge

[Ask Hjorth Larsen]

Hmmm.

Unfortunately spin-polarized calculations are always unreasonably
tricky.  Nobody knows why.

Nevertheless the calculation can be made to converge with some
violence to healthy physical principles.

Now it uses the dzp basis (any reason to use the default low-quality
basis set?).  It also uses the way too high electronic temperature
(0.2).

There are many more changes which are probably insignificant, although
it's generally a good idea to set initial magnetic moments to help
break spin symmetry, rattle the positions a bit to break any
geometrical symmetry, and getting rid of the nasty semicore Mn setup.

The below is mostly speculation as I didn't care to actually print out
and check what the electronic levels do, but it's probably not
completely wrong either.

Near the Fermi level there are probably several states, some of which
have different spins.  From one iteration to the next, the energies of
these states might switch around a bit.  This causes a "lot" (when you
want things to converge, even a little is a lot) of charge to move
from one spin to the other, completely messing up the XC potential.
This is worse than in non-spin-polarized calculations because there
are simply more energies, and I imagine that the messing up of the XC
potential is somehow causing more confusion than the mere messing up
of the Hartree potential which at the very least is equally felt by
all states.

To converge the calculation one needs a high electronic temperature or
a very defensive mixing (low mixing parameter, large long-range
oscillation damping).  The idea of the defensive mixing is to prevent
the density from changing too much, so that energies of different
states do not change order particularly if they belong to different
spins.  Unfortunately the defensive mixing did not solve the problems
of levels jumping around uncontrollably.  Therefore I had to increase
the Fermi smearing to at least 0.2.  0.15 won't do it.  Whether or not
that smearing is reasonable for your purposes is another question.  It
is possible, once the system has been relaxed and things are nice,
that a lower smearing suffices to converge the electronic structure.

A related problem is that if one has an electron in one state, then
there might be another state which is empty and has a lower energy.
Then the electron must be moved to that other state.  This causes the
previous state to get a lower energy, and the electron must be moved
back.  And thus it can never converge unless one allows fractional
occupations, considered unphysical in molecules by some people.  In
the limit of low temperature, the Most Density-Functionally Correct
Solution appears to be for those two (or more!) states to be exactly
degenerate but with different occupations.  Being thus unable to avoid
fractional occupations in the first place, there is no justification
why a high temperature would be all that horrible.

What I have never understood is how other codes deal with this
problem.  Maybe the mixers move charge between the two spins in a more
controlled manner, and are lucky enough to not actually have the
problem of states switching around.

http://dcwww.camd.dtu.dk/~askhl/files/mpc.py

Best regards
Ask

2014-11-26 12:40 GMT+01:00 Torsten Hahn <torstenhahn at fastmail.fm>:
>
>
>> Am 26.11.2014 um 12:20 schrieb Ask Hjorth Larsen <asklarsen at gmail.com>:
>>
>> Hello
>>
>> Could you please attach the geometry in a secure format (cube, xyz etc. or gpaw text output)? The traj format allows arbitrary code execution.
>>
>> Best regards
>> Ask
>>
>
>
> Ooops, i never noticed that.
>
> I attached it as *cube.
>
> Best,
> Torsten.
>
>
>> El 26/11/2014 11:56, "Torsten Hahn" <torstenhahn at fastmail.fm> escribió:
>>
>> Hi,
>>
>> im having difficulties to converge this simple molecular structure (spin polarized, lcao-mode, … see mpc.py). I already played around with different Mixer settings / occupations etc. The molecule converges fine in other DFT codes (NRLMOL).
>>
>> Sometimes a small number of geometry-optimization cycles runs with success, but sooner or later the calculation dies with KohnShamConvergence - Error.
>>
>> I realized that adding a hubbard_U > 3eV to Cobalt makes the calculation „stable“ in the sense that i see reliable convergence of the SCF cycle. Is it possible that there is a problem with the Co-Setup (Version 0.9.11271 ) ?
>>
>>
>> Best,
>> Torsten.
>>
>>
>>
>>
>> _______________________________________________
>> gpaw-users mailing list
>> gpaw-users at listserv.fysik.dtu.dk
>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users