The ABINIT Users Mailing List ( CLOSED )

["Anglade Pierre-Matthieu" <anglade@gmail.com>]

Hi,

Your procedure looks very natural. Yet beware. If not converged enough
your 'unconverged' calculation won't get you anycloser to the exact
solution. There has been some testimony on the forum of such problems
with your approach.

Depending on your system nature the problem complexity can increase or
not and so the number of steps. With no longue range coupling,
converging the 50 atoms system is not more difficult than converging a
200 atoms one. On the other hand with difficult systems N time the
number of atoms may mean N**2 time the number of iterations.

regards

PMA


On 4/4/07, Chol-Jun Yu <yucj@ghi.rwth-aachen.de> wrote:
> Dear users,
>
> thank you very much for your recommendations. In fact, our work is about very
> complicated system, that is, vicinal surface with kink (about 170 atoms).
> This situation makes the time of relaxation very crucial. I can't predict
> time of atomic relaxation (100 atoms are allowed to move). Can anyone predict
> the typical time?
>
> In order to save the time for relaxation, I thoght the following procedure.
> Since we don't know exact coordinates of atoms at the beginning of the
> simulation, we usually approximate the initial configuration. If this
> configuration is far away from the exact (within allowable error) one, the
> relaxation will take very much time, which is the most time-consuming of the
> work.
>
> While so as to obtain the sufficient reasonable result, we should use enough
> large parameters such as cutoff energy, special k-points, lattice constants
> of the supercell. We should check whether our parameters can guarantee the
> reasonable results or not. But it is almost not possible in the case of very
> time-consuming work.
>
> So I think that we can perform the relaxation of the atoms by the use of small
> ecut and only one k-point, so that we can get some good estimation for the
> initial configuration within short time. Then we can use this configuration
> for further relaxation with increased parameters, which maybe will finished
> within not so many Broyden steps. Finally we will check our results according
> to the paramters through the total energy calculation, with relaxed
> configuration. Does anyone criticize about this procedure?
>
> With best regards,
>
> Yu.
>
> On Wednesday, 4. April 2007 17:45, Michel Côté  wrote:
> > Actually, 60 steps for 50 atoms is not so bad. You have 50*3 degrees of
> > freedom and you get it relaxed in much less than that number. Remember that
> > when relaxing a structure, the values of the forces need to be precise
> > enough for the convergence, this means that you should use toldff for
> > criteria and use a value less than 5d-5.
> >
> > Michel
> >
> > Le 04/04/07 08:31, «Anglade Pierre-Matthieu» <anglade@gmail.com> a écrit:
> > > Hi,
> > >
> > > Apparently there is no better value for ionmov at present. I don't
> > > know the reasons why this algorithm is supposed to be most efficient
> > > for system with few degrees of reedom either. This last statement
> > > seems surprising to me since Broyden method is a very good algorithm
> > > for local minimisation. If you really need it you might try to
> > > implement a preconditioner to your minimisation problem. This may be
> > > the only efficient improvement you can add to BFGS. It consist in
> > > adding an explicit hessian guess to the implicit hessian evalutaion
> > > done by BFGS.
> > >
> > > Regards
> > >
> > > PMA
> > >
> > > On 4/4/07, Chol-Jun Yu <yucj@ghi.rwth-aachen.de> wrote:
> > >> > Dear users,
> > >> >
> > >> > I found that it could take much time (sometimes BFGS steps more than
> > >> > 60 in order to arrive convergence, 5*10-5 Ha/Bohr) to relax the atomic
> > >> > positions
> > >>
> > >> if
> > >>
> > >> > the number of atoms is much larger more than 50, for example, for
> > >> > surface supercell. I usually used the ionmove=2. Is it possible to
> > >> > reduce the BFGS steps by using other values of "ionmov" parameter?
> > >> >
> > >> > With best wishes,
> > >> >
> > >> > Yu Chol Jun.


--
Pierre-Matthieu Anglade