[Wien] SCF convergence velocity

[Peter Blaha]

> First of all I want to thank you for all previous answers to my questions.
> Today I would like to ask about velocity of scf cycles. I'm calculating
> 88 atoms big system (Fe on W) and after something like 150 iterations,
> the charge convergence parameter equals to ~0.05 and it converges very
> slowly.
> I attached(I hope it is allowed) my struct file and a plot of charge
> convergence versus number of iteration. One can see that during ca 70
> iterations
> starting from 60 till 130 the parameter stays almost constant.
>
> I' doing spin polirized calculations, the RMT radius I've chosen is 2.45
> , I also reduced mixer factor to 0.01.

Unfortunately some systems (in particular magnetic ones) converge very
slowly.

Looking at your curves:
DIS came down quite reasonable for the first 40 iterations. There are no
oszillations observable. From this observation alone I would expect rather
smooth convergence and you are simply "overdamped".

After that it does not really proceed (except some decrease from 60-70 and
at the very end (140-)

Could it be, that you damped the convergence too much by using such a small
mixing factor (0.01 + BROYD ?)? It is sometimes necessary to increase it
slowly once a certain convergence has been reached. Did you try a larger
mixing ? Did it diverge ? At what :DIS ?

General strategy:

"save" your results every 20 iteration. This way you can always "recover"
in case of divergence when mixing has been increased.

Check the "real" mixing factor printed in case.scf. mixer reduces this factor
automatically depending on :DIS and the number/type of atoms.

Increase your mixing factor slowly (every 20 BROYD cycles) until you see
fluctuations coming up. Check not just DIS, but find with DTO which atoms
cause problems, and then with  NTOxxx (or NUP/DN) compared to CTOxx whether
NTOxxx oszilates or is always to small/large. In the latter case increase
mixing.

If NTOxxx oszillates, 2 improvements are possible:

Sometimes a better k-mesh can solve those problems.

Sometimes a smearing method (like TEMP or GAUSS) helps, but be aware,
that too large smearing parameter could in particular whith magnetic
systems even lead to nonmagnetic solutions (Nevertheless it is worth
a trial like TEMP 0.010)

Regards
                                      P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------