The ABINIT Users Mailing List ( CLOSED )

[Xavier Gonze <xavier.gonze@uclouvain.be>]

Dear Hichem Bouderda,

On 10 Oct 2009, at 15:12, hichem.bouderba@gmail.com wrote:

> hello
> first :
> in the hcp structure to sample the reciprocal space (k grid) I  
> propose two
> examples:
> 1 - I specify	ngkpt 6x6x8 (just an example),	I get nkpt=28 and  
> kptrlen =
> 42.9
> 2 - I specify kptrlen = 46 (without specifying ngkpt nor  
> kptrlatt) ,  I get
> nkpt= 16
> is the following correct :
> nkpt2 < nkpt1 => the second grid is faster.
> kptrlen2 > kptrlen1 => the second grid is more accurate too.

Right. See the  doc related to the shiftk input variable for a comment  
about sampling the BZ for the HCP structure
> second:
>
> here is a typical comment on k-point sampling in the litterature:
> " Brillouin zone integrations were performed using
> 4000/(number of atoms in unit cell) k-points distributed
> as uniformly as possible on a Monkhorst-Pack mesh."
> Y. Wang et al. / Computer Coupling of Phase Diagrams and  
> Thermochemistry 28
> (2004) 79-90
>
> this is an indication of the sampling accuracy. so one would expect  
> that for
> different structures this sampling would give the same level of  
> accuracy for
> integration over the brillouin zone.
> question: why one would care about the sampling way if the results are
> converged for the studied physical quantity. May be for some  
> structure I would
> be converged for 10x10x10 and for an other one I would wait 16x16x16  
> sampling
> to get convergence and the converged values are equivalent in term  
> of accuracy.

Right. But you can use a rule of thumb.
"4000/(number of atoms in unit cell)"  seems to me to lead to an  
accurate sampling,
needed typically for metals with a high DOS at Fermi level.

Xavier
> comments, "warnings" are welcome !