[CPMD-list] How to pick a good CUTOFF to run BO-MD as fast as it can?

[Axel Kohlmeyer]

dear rex,

On 5/31/06, osirixu-cpmd at yahoo.com.cn <osirixu-cpmd at yahoo.com.cn> wrote:
> Dear cpmders,
>
>      As we know, when the CUTOFF is too big, the md
> calculation will obviously be slower; but while the
> CUTOFF is too small, the convergence to the minimum
> energy will be slower, thus the whole md calculation
> will be slower, too. So my question is: what pre-test

as alex already pointed out, the plane wave cutoff (i.e. the size of your
basis set) has to be chosen in a way, that all relevant properties are
converged within required accuracy. for MD the property that matters
are the forces (for variable cell MD the stress tensor, btw). so you have
to choose your plane wave cutoff so, that the forces are accurate enough.

for BO-MD you also have to pay attention to the CONVERGENCE ORBITALS
parameter which also has an influence on the accuracy of the forces.
to have reasonable energy conservation, you have to use a tighter
convergence parameter, about 1.e-6 to 1.e-7.

> can we do to pick the best CUTOFF to make the whole
> BO-MD claculation fastest?

in CPMD there are several options that can make your
BO-MD faster most of them at the expense of increased
memory usage.

a) the selection of the wavefunction optimizer. the default ODIIS 10
is usually fastest for 'well behaving' systems. sometimes it helps to
reduce to ODIIS 5 and add NO_RESET=50. in case of less well behaving
systems, you may be better off using PCG MINIMIZE (or plain PCG).

b) with REAL SPACE WFN KEEP you reduce the number of  FFTs. this helps
especially on parallel jobs with high latency networks. this will crank up the
memory use significantly. you can specify a limit and cache only a part of
the FFT (see manual).

c) with EXTRAPOLATE WFN on top of that you can improve the initial guess
for subsequent wavefunction optimizations by extrapolating it from previous
wavefunctions. using a 3rd or 4th order extrapolation seems to work very well
and roughly doubles the memory requirements on top of REAL SPACE WFN KEEP.

of course during the equilibration phase of your simulation, you can reduce
the plane wave cutoff a little bit to sacrifice a little(!) accuracy
for more speed,
but as soon as you are close to equilibrium, you should go to the final value.

having a fast MD with bogus forces is a waste of CPU time.

regards,
     axel.
>
>      Any advice or suggestions are greatly
> appreciated!
>
>      Many Thanks!
>
>       Rex
>      May 31, 2006
>
>
>
>
>
>
> ___________________________________________________________
> Mp3·è¿ñËÑ-иèÈȸè¸ßËÙÏÂ
> http://music.yahoo.com.cn/?source=mail_mailbox_footer
> _______________________________________________
> CPMD-list mailing list
> CPMD-list at cpmd.org
> http://cpmd.org/mailman/listinfo/cpmd-list
>
>


-- 
=======================================================================
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.