On 11/03/2015 02:47 PM, Laurence Marks wrote:
> a) :FCHECK (bottom of case.scf) was large. This is the sum of all
> the forces, and should be small. Particularly for cells without
> inversion one can get bad, highly asymmetric densities in which case
> MSR1a can have problems.

True, I had not notice this before.  :FCHECK was already quite large
(hundreds for x and y, even >1000 for z) before ‘clminter’ and “jumped”
to even larger values afterwards.

> b) The greed is small. Too small a value can be as bad as too large.
> I have struggled with this for years and failed to find a strong
> ansatz for this, although I believe the next release of the mixer
> will be better.

AFAIK, the greed in MSR1(a) is set internally by the mixer (the
corresponding value in case.inm being ignored), so there is nothing I
can do directly to influence this, is there?

> I am not sure what the calculation is, perhaps an oxide surface
> where you have made a guess at the initial structure and want to
> minimize to something more reasonable.

Quite a good guess, it is an oxide heterostructure including a slab of
vacuum, but the initial structure is derived (cut out) from a converged
one from an older calculation, so I would have expected only relatively
minor adjustments.  As such, the large forces also come as something of
a surprise.

> I strongly suggest trying to use cells with inversion, they behave
> much, much better.

In this case, inversion symmetry could only be achieved by adding an
additional “film” on the back side of the “substrate”.

> I also strongly suggest that you look at the Bond Valence Sums (BVS)
> and tweak the initial positions until they are reasonable. (x nn ; 
> grep Bond *tnn). If, for instance, you have highly underbonded O
> (e.g. 0.8) it can take forever and the calculations can be unstable
> -- convergence is faster the more physical are the atomic positions.

I guess the expectation is that the BVS should be close to the “formal
valence” of the ion, right?  In this case, it seems okay: The first BVS
value is between 1.4 and 2.2 for O, deviations for other species are
rather smaller).

> Good luck.

Thanks, and thank you for your tips.


Elias Assmann
Institute of Theoretical and Computational Physics
TU Graz                   ⟨https://itp.tugraz.at/⟩

Attachment: signature.asc
Description: OpenPGP digital signature

Wien mailing list

Reply via email to