Geometry optimization convergence problem

[Peter]

Dear all,

I'm trying to optimize geometry of a large (~13000 atoms, it's a
protein) mechanical (FIST using charmm27 FF) isolated system.
Unfortunately both CG and LBFGS fails to converge to reasonable values
of maximum gradient (4.5E-4 Hartree/Bohr). After approximately 1e3
steps, maximum gradient value starts to oscillate around 2e-3 and
leaves no hope for convergence (see fig.
http://groups.google.com/group/cp2k/web/max_grad_vs_step.gif).

How could I tune either of the algorithms to achieve convergence?

I've uploaded archive with my input, structure and FF parameters files
here:
http://groups.google.com/group/cp2k/web/geo_opt-no_convergence.tar.gz

Thanks in advance,
Peter

P.S. Here are the contents of one of the input files (LBFGS), so that
you needn't to untar the archive to take a look at it:

&GLOBAL
  PROJECT OPT-LBFGS
  RUN_TYPE GEO_OPT
  PREFERRED_FFT_LIBRARY FFTMKL
&END GLOBAL

&MOTION
  &GEO_OPT
    TYPE MINIMIZATION
    MAX_ITER 20000
    OPTIMIZER LBFGS
    &LBFGS
      MAX_H_RANK 10
      MAX_F_PER_ITER 100
    &END LBFGS
  &END GEO_OPT

  &PRINT
    &TRAJECTORY
      FORMAT XYZ
      FILENAME =lbfgs_opt.xyz
    &END TRAJECTORY
  &END PRINT
&END MOTION

&FORCE_EVAL
  METHOD FIST
  &MM
    &FORCEFIELD
      PARM_FILE_NAME par_all27_prot_lipid.prm
      PARMTYPE CHM
    &END FORCEFIELD

    &POISSON
      PERIODIC NONE
      PSOLVER WAVELET
      &EWALD
        EWALD_TYPE NONE
      &END EWALD
    &END POISSON

    &NEIGHBOR_LISTS
      GEO_CHECK F
    &END NEIGHBOR_LISTS
  &END MM

  &SUBSYS
    &CELL
      ABC 100 100 100
      PERIODIC NONE
    &END CELL

    &TOPOLOGY
      COORD_FILE_NAME 1AIG-apo_fe_uq_wat.pdb
      COORD_FILE_FORMAT PDB
      CONN_FILE_NAME 1AIG-apo_fe_uq_wat.psf
      CONN_FILE_FORMAT PSF
      CENTER_COORDINATES
    &END TOPOLOGY
  &END SUBSYS
&END FORCE_EVAL