[Pw_forum] convergence NOT achieved after 100 iterations --- related question

Hande Ustunel hande at newton.physics.metu.edu.tr
Fri Apr 27 11:38:11 CEST 2012


Dear Nicola,

Many thanks for the prompt reply.

> first question - would mp work?

This we haven't tried. Since mv and mp give very similar numerical results
for a given smearing width I guess I thought I should try something a bit
different, i.e. gauss.

> In terms of "structure" of the smearing function, gaussian is
> the smoothest, then mv, then mp. So if there were an intrinsic problem
> with states at the fermi energy, it should be even more apparent in
> mp .

I see. I'll surely take a look at mp then.

> As a general comment, the role of smearing is to help improve the
> accuracy of k-point sampling - and so the 0.01-.02 you are using
> is exactly in the right ball park.

Sure. Actually using a larger kpoint set sometimes helps but it gets a bit
tricky because if you've run a number of calculations with a smaller kpoint
set (say for a publication), then you want to stick to that. So sometimes
we actually try with a larger set and then once (if) we see convergence
revert back to the one we want. This only works about half the time
though. Our g-Al2O3 + precious metal atom calculations were particularly
notorious so we got to try this close to about 20-30 times.

> A side effect (a positive one) for smearing is to help convergence
> in iterative algorithms (as in PWSCF) - for that side effect,
> 0.01-0.02 is actually small, and so the more "structure" of
> mv or mp could hamper the iterative convergence. So you could try
> converging first with a larger smearing (.04 to .08), and then restart
> with the "right" one. I'd be somewhat surprised to see much difference
> between gaussian/mp/mp for say 0.07 Ry ~ 1 eV of smearing. Of
> course, such a large smearing shoudl be used only to help
> get convergence, and then reduced.

Oh, wow. We haven't done that since I'm a bit scared of large smearings but
this is a great suggestion. We do make use of a similar trick though,
namely going back and forth between an initial gauss shot followed by mv
but I'm afraid this works even less than half the time. But I will
definitely give this a shot and report back the result.

Many thanks once again.
Best wishes,
Hande

--
Hande Toffoli
Department of Physics
Office 439
Middle East Technical University
Ankara 06531, Turkey
Tel : +90 312 210 3264
http://www.physics.metu.edu.tr/~hande



More information about the users mailing list