[Pw_forum] Iron cluster relaxation

[Axel Kohlmeyer]

On Mon, 13 Nov 2006, Cyrille Barreteau wrote:

CB> In fact I have looked in more details at the results of pw.x calculations
CB> and there is a funny behaviour. The restarted calculations converges towards
CB> a totally different scf solution with a total magnetization  1mu_B larger
CB> than the previous "crashed" calculation. I guess there are several magnetic
CB> solutions and I "jumped" from one solution to the other.. Funny..

hmmmm.... interesting. that could explain the geometry optimizer 
getting 'stuck' even if you use a large enough cutoff and tight
enough wavefunction convergence. since you always restart from
the previous wavefunction, there is a high tendency to stick to
a similar state, even if it is not the optimal one anymore. 
did you restart from the previous wavefunction/density
or with a new (atomic or random) wavefunction?

sadly, this kind of situation is one of the problems that can happen,
sometimes rather unexpectedly (it happened to me in a CP dynamics
only after several ps simulation time when the system seemed 
equilibrated and i switched the thermostats for production run).

how about adding a (paranioa) flag to pw.x that allows to do a final 
wavefunction optimization from an initial wfc (either random or 
atomic) at the last step of a geometry optimization and then 
compare the energies/eigenvalues and warn if the differences
are large (the flag could be a floating point number with 0.0
meaning no check)?

cheers,
   axel.


CB> 
CB>   cyrille
CB> 
CB> 
CB> > Dear Kostya and Axel
CB> >
CB> > I have restarted my job with the last coordinates and
CB> > the geometry optimization converged in just one scf cycle!!
CB> >
CB> > ==========================================================
CB> >     BFGS Geometry Optimization
CB> >
CB> >     bfgs converged in   1 scf cycles and   0 bfgs steps
CB> >
CB> >     End of BFGS Geometry Optimization
CB> > =============================================
CB> > So I guess the previous "crashed run" did converge correctely
CB> > toward the right optimal atomic configuration!!
CB> >
CB> > thanks
CB> >
CB> >    cyrille
CB> >
CB> >> On Fri, 10 Nov 2006, Konstantin Kudin wrote:
CB> >>
CB> >> kostya,
CB> >>
CB> >> KK>  It is not clear why the job should have a hard crash like this 
CB> >> instead
CB> >> KK> of either terminating gracefully, recovering from the condition 
CB> >> in some
CB> >> KK> way, or at least printing the very thing that you say, i.e. that
CB> >> KK> "optimization terminated due to noise in the forces, increase 
CB> >> cutoff",
CB> >> KK> or something else similar.
CB> >>
CB> >> i agree. after this issue has come up quite a few times in the last 
CB> >> months, it may be worth putting in the effort
CB> >> and changing the code, so that it a) exits gracefully,
CB> >> b) has a few more options on how to handle the resets,
CB> >> e.g. force a minimum number of steps between two resets,
CB> >> or have a small(!) randomization of coordinates.
CB> >>
CB> >> that just leaves the usual problem: who is going to do it?
CB> >> anybody in the market for making a lot of people happy?
CB> >>
CB> >> KK>  Actually, in my experience just restarting the job with the last
CB> >> KK> coordinates has a very good chance to advance further, and even 
CB> >> find a
CB> >> KK> lower energy structure since the BFGS will be started from scratch.
CB> >>
CB> >> hmmm. if you see the latter, then perhaps the reset is not
CB> >> a complete reset and there may be a bug (unless you have
CB> >> a 'randomization' of the coordinates due to limited
CB> >> accuracy when reading in the so far optimized coordinates).
CB> >>
CB> >> cheers,
CB> >>   axel.
CB> >>
CB> >> KK> KK>  Kostya
CB> >> KK> KK> KK> KK> --- Axel Kohlmeyer <akohlmey at cmm.chem.upenn.edu> wrote:
CB> >> KK> KK> > On Fri, 10 Nov 2006, Cyrille Barreteau wrote:
CB> >> KK> > KK> > CB> Dear Pwscf users
CB> >> KK> > KK> > dear cyrille,
CB> >> KK> > KK> > did you check the mailing list archives?
CB> >> KK> > KK> > IIRC, this error seems to happen if you try to
CB> >> KK> > optimize the geometry too tightly for the
CB> >> KK> > given convergence of the wavefunction or
CB> >> KK> > when you suffer from 'ripples' due to a
CB> >> KK> > too small density cutoff.
CB> >> KK> > KK> > salut,
CB> >> KK> >   axel.
CB> >> KK> > KK> > CB> I am trying to relax an iron cluster (cubocathedron) and
CB> >> KK> > CB> pw.x systematicaly ends with the following error:
CB> >> KK> > CB> KK> > CB> KK> >
CB> >> KK> 
CB> >> %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 
CB> >>
CB> >> KK> > CB>      from bfgs : error #         1
CB> >> KK> > CB>      bfgs history already reset at previous step
CB> >> KK> > CB> KK> >
CB> >> KK> 
CB> >> %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 
CB> >>
CB> >> KK> > CB> KK> > CB>      stopping ...
CB> >> KK> > CB> 2
CB> >> KK> > CB> KK> > CB> KK> > CB> You can find the input and output files 
CB> >> on the following web
CB> >> KK> > site:
CB> >> KK> > CB> KK> > CB>
CB> >> KK> >
CB> >> KK> 
CB> >> http://www-drecam.cea.fr/Images/Pisp/cbarreteau/temporaire/cuboctahedron_relax.in 
CB> >>
CB> >> KK> > CB>
CB> >> KK> >
CB> >> KK> 
CB> >> http://www-drecam.cea.fr/Images/Pisp/cbarreteau/temporaire/cuboctahedron_relax.out 
CB> >>
CB> >> KK> > CB> KK> > CB> KK> > CB> thanks in advance for any constructive 
CB> >> comment :-)
CB> >> KK> > CB> KK> > CB>    cyrille
CB> >> KK> > CB> KK> > CB> KK> > KK> > -- KK> >
CB> >> KK> 
CB> >> =======================================================================
CB> >> KK> > Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu  KK> > 
CB> >> http://www.cmm.upenn.edu
CB> >> KK> >    Center for Molecular Modeling   --   University of Pennsylvania
CB> >> KK> > Department of Chemistry, 231 S.34th Street, Philadelphia, PA
CB> >> KK> > 19104-6323
CB> >> KK> > tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel:
CB> >> KK> > 1-215-898-5425
CB> >> KK> >
CB> >> KK> 
CB> >> =======================================================================
CB> >> KK> > If you make something idiot-proof, the universe creates a better
CB> >> KK> > idiot.
CB> >> KK> > KK> > _______________________________________________
CB> >> KK> > Pw_forum mailing list
CB> >> KK> > Pw_forum at pwscf.org
CB> >> KK> > http://www.democritos.it/mailman/listinfo/pw_forum
CB> >> KK> > KK> KK> KK> KK>  KK> 
CB> >> ____________________________________________________________________________________ 
CB> >>
CB> >> KK> Cheap talk?
CB> >> KK> Check out Yahoo! Messenger's low PC-to-Phone call rates.
CB> >> KK> http://voice.yahoo.com
CB> >> KK> _______________________________________________
CB> >> KK> Pw_forum mailing list
CB> >> KK> Pw_forum at pwscf.org
CB> >> KK> http://www.democritos.it/mailman/listinfo/pw_forum
CB> >> KK>
CB> >>  
CB> >>
CB> >
CB> >
CB> 
CB> 
CB> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.