[Wien] nHOw to accelerate the caculation?

[Peter Blaha]

> For some time, we are using the strategy of relaxing large unit cells with
> vasp, and then doing one static calculation with wien afterwards (to obtain
> hyperfine field information). While this is definitely much faster than
> relaxing with wien (often an order of magnitude), it is not clear how
> reliable the results are. In many cases the wien-forces in the vasp-relaxed
> structure are quite small (< 5 mRy/au), while in others they can be up to 20
> or 30 mRy/au. That seems to be related to the elements involved (and

It is clear (and it always will be like that), that an FP-APW+lo method
will never be the fastest code. However, for this drawback you gain that
you are working with a very reliable method, i.e. one which works for all
elements and compounds. When using any other code (either PP or LCAO type
methods) you will always first have to test this against WIEN - and maybe
then you can use it (and be faster).

In addition there are classes of compounds, where APW will always be quite
slow. This happens when you are forced to use small spheres because of
small nn-distances. (Examples are in particular organic materials with
short C=O,... bonds)

Structural relaxation is unfortunately far from being optimized in WIEN.
(We are working on that).
a) You need to find out yourself good parameters for the case.inM file.
Setting the delta's too small may result in many timesteps, setting them
too large it may even oszillate, never converge (ie. diverge).

It has been described several times what you need to do:
Monitor   grep :FORxxx *scf_mini
When the forces oszillate (change sign), the delta's must be reduced, when
they change too little, deltas should be increased.

In addition, several other steps may help speeding up things:

Suppose at the beginning the forces are big: i.e. 20-50 mRy/bohr.

Then reduce    the k-mesh
     and       RKmax
such that the resulting forces are "qualitatively correct"
this means: it does not matter whether the force is 50 or 40 mRy/bohr,
it will move the atom in the proper direction anyway.

In addition, it is not necessary to converge the scf cycle perfectly well,
i.e.    run_lapw -fc 5      might be good enough.

With these reduced RKMAX, k-mesh and scf-cycles it should be pretty fast
and I would not expect that PP methods are significantly faster.

Once all forces are below e.g. 5-10 mRy/bohr, increase all parameters
(maybe you can also increase the "delta's" again).

You could put this strategy into a little shell script (copying the
more/less accurate input files to the actual used one) to make it more
convinient. However, you still have to monitor :FORxxx in *mini and
adjust the "delta's" by hand.

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/
--------------------------------------------------------------------------