OUTER_SCF

[marc]

Dear all,

I'm trying to get a better understanding of the different options that
are available in cp2k by reading some papers and also to get a more
hands-on experience with DFT-calculations I'm comparing the results of
different functionals. To get some help I was reading the cp2k user
self-support(part 3) and there I came across the OUTER_SCF option
(FORCE_EVAL%DFT%SCF%OUTER_SCF). So my question
is what does an outer_scf cycle do exactly? I didn't find any
reference to the literature or something similar... Apparently it is
important for the convergence but I don't really see how. And what is
the relation to the 'inner scf' cycle? There should be an important
interaction between the two because the total number of iteration
steps is the product of the maximum internal and maximum external
steps...
Maybe this explained somewhere in a paper?

Thanks in advance,

marc

[Jörg Saßmannshausen]

Dear Marc,

as far as I understood it from the Tutorial, it works like that:
Imagine 2 loops, the inner and the outer SCF cycle. First you run the inner
cycle. If it converges, you only have one outer cycle done (starting the
inner cycle).
If the inner cycle does not converge, rather then going down 'the wrong path',
you go back (into the outer cycle), make some corrections, and go into the
inner cycle again (so second outer circle run). By doing so, the chances of
convergence are increased.
I am sure somebody an give a less pictorial and more mathematical
interpretation of that, but that was my understanding at the time.

I am aware that this methode is used by other programs as well but in a
different manner.

Right, I am tired, braindead and prolly that is full of typos.

All the best

Jörg

--
*************************************************************
Jörg Saßmannshausen
Research Fellow
University of Strathclyde
Department of Pure and Applied Chemistry
295 Cathedral St.
Glasgow
G1 1XL

email: jorg.sassm...@strath.ac.uk
web: http://sassy.formativ.net

Please avoid sending me Word or PowerPoint attachments.
See http://www.gnu.org/philosophy/no-word-attachments.html

[Fawzi Mohamed]

On 3-apr-09, at 01:36, Jörg Saßmannshausen wrote:

>
> Dear Marc,
>
> as far as I understood it from the Tutorial, it works like that:
> Imagine 2 loops, the inner and the outer SCF cycle. First you run
> the inner
> cycle. If it converges, you only have one outer cycle done (starting
> the
> inner cycle).
> If the inner cycle does not converge, rather then going down 'the
> wrong path',

Not down the wrong path, but simply converging slowly

> you go back (into the outer cycle), make some corrections,

This "correction", if you use the outer cycle just to improve the
convergence is simply recalculating the preconditioner of OT.

The preconditioner might depend on the actual C (this is not true for
example for the kinetic one), and in this case a better C means a
better preconditoner, which means better convergence.

As building the preconditioner costs time (especially for the better
ones like full_all) you do not want to rebuild it each step, but with
the outer SCF you can say that if you don't converge in MAX_SCF
(inner) steps, then the preconditioner should be rebuilt, and you
should try again (hopefully with better convergence), repeating this
procedure at most (outer) MAX_SCF times.

As consequence when used in this way you should have the same EPS_SCF
in both cycles.

Fawzi

[flo]

Just a short remark. The outer loop optimization is used as well if
you apply constraints on the density (DDAPC,S2,...). In this case the
outer scf can have a different convergence criterium, since the
variable to optimize is a different one than the SCF convergence
criterium.

Flo

[marc]

OK I understand.

Thanks very much guys

marc