[CPMD-list] variable cell optimization troubles

[Axel Kohlmeyer]

On Sun, 2 Mar 2008, radbalu at comcast.net wrote:

RB> Dear All,

dear rad,

RB> I am trying a variable cell optimization of a molecular crystal 
RB> using DCACP based pseudopotentials. when I used regular TM psuedos 
RB> without the dispersion the geometry converged. However when I used 

please be more precise, are you using the goedecker format DCACPs
or the numerical ones with the additional projector?
also which version of CPMD, what platform etc. etc.

RB> DCACP terms and the following input I have problems getting 
RB> convergence:

RB> &CPMD
RB>  OPTIMIZE GEOMETRY
RB>  STRESS TENSOR
RB>      5
RB>  PARRINELLO-RAHMAN
RB>  STEEPEST DESCENT  CELL
RB>  PRINT FORCES ON
RB>   PCG MINIMIZE NOPRECONDITIONING
RB>   TIMESTEP
RB>    10
RB>   EMASS
RB>    600
RB>  CONJUGATE GRADIENTS IONS
RB> &END
RB> 
RB> &SYSTEM
RB>  ANGSTROM
RB>  SCALE
RB>  CELL VECTORS
RB> 

RB> "** STRESS| SIZE OF THE PROGRAM IS 257652/24861248 kBYTES ***" gets 
RB> printed implying the stress tensor is calculated. Then there is a 
RB> discontinuity in the energy and the optimization proceeds in a 
RB> different direction without ever converging.

these kind of discontinuities usually point to memory corruption.
which in turn can have a lot of reasons. your input is very complex
and thus strange interactions between different modules are possible.

[...]

RB> Could some one help understand what is going on and point out 
RB> anything that is odd in the way I set up the input?

i have serious doubts if anybody has tried or tested the kind of
calculation that you are trying (with CPMD). there are a lot of
possibly contributing factors (stress tensor with goedecker vs. 
numerical), variable cell geometry optimization (why not simulated
annealing? or a set of geometry optimizations at varying but fixed
cell dimensions?).

as has been said here many times, CPMD offers so many options that 
it is impossible to check for all combinations. as a consequence 
your chances for a specific combination of flags to work and work
well increase the more commonly a feature is used and decrease 
the more exotic your input is.

RB> Thank you in advance for your suggestions.

without sufficient information and a complete input one can 
only speculate. your best shot is to dig into the sources,
look at the flow of code and come up with test cases for
each feature being used and if you find something odd and
try to correct it (and let us through, e.g, cpmd at cpmd.org 
know what should be changed). 

cheers,
   axel.

RB> 
RB> Rad
RB> _______________________________________________
RB> CPMD-list mailing list
RB> CPMD-list at cpmd.org
RB> http://cpmd.org/mailman/listinfo/cpmd-list
RB> 

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