[CPMD-list] Convergence problem with USPP

Tue Apr 1 19:05:01 CEST 2008

On Tue, 1 Apr 2008, Linas Vilciauskas wrote:

linas,

LV> Thanks for the prompt reply and sorry for not providing enough
LV> details... I certainly followed the rules in the manual for selecting
LV> the box size for an isolated system calculation and at lest for the
LV> norm-conserving PPs the box size (in this case 8.0 A) was never a
LV> problem (neither with LDA nor GGA nor Hybrid XC functionals). Here is

well, the specifications for all poisson solvers state,
that the charge density has to be zero at the box boundary
and as a rule of thumb for the hockney solver you should
have minmax of the coordinates plus 3 angstrom.
if i dump the electron density with RHOOUT and look
at it, you have even electron density crossing "through"
the periodic images. if i compute the min/max of your 
coordinates plus 3 angstrom i end up with a 9.5x9.5x7.2 AA^3 box.
...and this would be the absolute minimum, not accounting
for any movement of the molecules during the optimization.

due to the associated higher level of noise in the vacuum
areas, low cutoff calculations are more sensitive to too
small boxes.

LV> my full input:

re-reading your original post, i noticed that i may have 
misunderstood your initial question. am i reading it right that 
it is _not_ the individual wavefunction optimizations that fail 
to converge (like i initially thought), but the geometry 
optimization themselves.

in that case you may also have become a victim of "ripples",
since the real space grid also has to accommodate the augmentation 
charges. setting DUAL to a value of 5.0 is most likely not sufficient,
particularly with a cutoff of 25ry which is a the lower end for
what is reasonable for USPPs. people who want to compute accurate
phonons for tricky systems, may need a wavefunction cutoff of 
40-50ry and a DUAL value of 8-10.

for accurate results (please also see my comments on the convergence 
of forces w.r.t. wavefunction convergence in a previous mail).

so i made two tests. using a box of 10x10x10, a CUTOFF of 35
a DUAL of 6, and CONVERGENCE ORBITALS of 5e-7, i get a converged
geometry in 12 steps. with a CUTOFF of 25 and DUAL of 5.0
convergence of the geometry is already a bit more tricky, i.e,
one can see the ripples effect and it takes 21 geometry steps
to converge. 

based on those observations, i would recommend to re-check some
of your previous calculations, if they were using the same coordinates
and box size. the artefacts introduced by the poisson solvers can be 
quite substatial in unfavorable circumstances. the fact that a
calculations converges or finishes normally does not indicate that 
it is completed correctly. :-(

cheers,
  axel.

LV> 
LV> Sample input:
LV> 
LV> &INFO
LV> &END
LV> 
LV> &CPMD
LV>   OPTIMIZE GEOMETRY XYZ
LV>   STRUCTURE BONDS ANGLES DIHEDRALS
LV>   PCG MINIMIZE
LV>   ODIIS NO_RESET=30
LV>    10
LV>   MAXSTEP
LV>    1000
LV> &END
LV> 
LV> &DFT
LV>   FUNCTIONAL BLYP
LV>   GC-CUTOFF
LV>    5.D-6
LV> &END
LV> 
LV> &SYSTEM
LV>   SYMMETRY
LV>    0
LV>   ISOLATED MOLECULE
LV>   ANGSTROM
LV>   CELL
LV>    8.0 1.0 1.0 0.0 0.0 0.0
LV>   CUTOFF
LV>    25.0
LV>   DUAL
LV>    5.0
LV> &END
LV> 
LV> &ATOMS
LV> *P_VDB_BLYP.psp FORMATTED
LV>   LMAX=P
LV>    2
LV>   -2.394917   -0.130691   -0.442475
LV>   -0.323809    2.480816    1.970286
LV> 
LV> *O_VDB_BLYP.psp FORMATTED
LV>   LMAX=P
LV>    8
LV>   -3.800575   -0.695687   -0.967861
LV>   -2.220024   -0.256766    1.033758
LV>   -1.358141   -0.988753   -1.307625
LV>   -2.282227    1.318659   -1.030921
LV>   -1.161599    1.418333    2.763236
LV>   -0.755639    2.744989    0.566853
LV>   -0.317814    3.792773    2.892051
LV>    1.188019    1.977302    2.114271
LV> 
LV> *H_VDB_BLYP.psp FORMATTED
LV>   LMAX=S
LV>    6
LV>   -4.543162   -0.409366   -0.423954
LV>   -1.227273   -1.878121   -0.959249
LV>   -1.684876    1.918836   -0.497156
LV>   -1.571156    0.717894    2.177084
LV>   -1.109816    4.330419    2.778487
LV>    1.800148    2.469017    1.554849
LV> &END
LV> 
LV> 
LV> On Tue, Apr 1, 2008 at 11:58 AM, Axel Kohlmeyer
LV> <akohlmey at cmm.chem.upenn.edu> wrote:
LV> > On Tue, 1 Apr 2008, Linas Vilciauskas wrote:
LV> >
LV> >  LV> Dear CPMD community,
LV> >
LV> >  dear linas,
LV> >
LV> >  [...]
LV> >
LV> >  LV> problem with USPPs for certain atoms/combinations? Does anybody have
LV> >  LV> some ideas?
LV> >
LV> >  i suspect your box is to small, but since you were omitting
LV> >  the most important part of the input to identify this,
LV> >  the coordinates, nobody can tell for sure. please have a look
LV> >  at the comments that the manual has about choosing the box
LV> >  for isolated system calculations.
LV> >
LV> >  cheers,
LV> >    axel.
LV> >  LV>
LV> >  LV> System info:
LV> >  LV> CPMD 3.11.1
LV> >  LV> OS: IBM AIX 5.2 L
LV> >  LV> FCC: IBM XL 8.1.1 (OpenMP only)
LV> >  LV>
LV> >  LV> Sample input:
LV> >  LV> &INFO
LV> >  LV> &END
LV> >  LV>
LV> >  LV> &CPMD
LV> >  LV>  OPTIMIZE GEOMETRY XYZ
LV> >  LV>  STRUCTURE BONDS ANGLES DIHEDRALS
LV> >  LV>  PCG MINIMIZE
LV> >  LV>  ODIIS NO_RESET=30
LV> >  LV>   10
LV> >  LV>  MAXSTEP
LV> >  LV>   1000
LV> >  LV> &END
LV> >  LV>
LV> >  LV> &DFT
LV> >  LV>  FUNCTIONAL BLYP
LV> >  LV>  GC-CUTOFF
LV> >  LV>   5.D-6
LV> >  LV> &END
LV> >  LV>
LV> >  LV> &SYSTEM
LV> >  LV>  SYMMETRY
LV> >  LV>   0
LV> >  LV>  ISOLATED MOLECULE
LV> >  LV>  ANGSTROM
LV> >  LV>  CELL
LV> >  LV>   8.0 1.0 1.0 0.0 0.0 0.0
LV> >  LV>  CUTOFF
LV> >  LV>   25.0
LV> >  LV>  DUAL
LV> >  LV>   5.0
LV> >  LV> &END
LV> >  LV>
LV> >  LV> &ATOMS
LV> >  LV> *P_VDB_BLYP.psp FORMATTED
LV> >  LV> ...............
LV> >  LV> *O_VDB_BLYP.psp FORMATTED
LV> >  LV> ...............
LV> >  LV> *H_VDB_BLYP.psp FORMATTED
LV> >  LV> ..............
LV> >  LV> &END
LV> >  LV>
LV> >  LV>
LV> >  LV>  --
LV> >  LV> Linas Vilciauskas
LV> >  LV> Max-Planck-Institut für Festkörperforschung
LV> >  LV> Heisenbergstraße 1
LV> >  LV> D-70569 Stuttgart, Germany
LV> >  LV> _______________________________________________
LV> >  LV> CPMD-list mailing list
LV> >  LV> CPMD-list at cpmd.org
LV> >  LV> http://cpmd.org/mailman/listinfo/cpmd-list
LV> >  LV>
LV> >
LV> >  --
LV> >  =======================================================================
LV> >  Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
LV> >    Center for Molecular Modeling   --   University of Pennsylvania
LV> >  Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
LV> >  tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
LV> >  =======================================================================
LV> >  If you make something idiot-proof, the universe creates a better idiot.
LV> _______________________________________________
LV> CPMD-list mailing list
LV> CPMD-list at cpmd.org
LV> http://cpmd.org/mailman/listinfo/cpmd-list
LV> 

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