Quickstep: SCF Energy increasing after few outer-SCF cycles

[Sharma SRK Chaitanya Yamijala]

Dear all CP2K users,

I am trying to optimize a Graphene Quantum Dot using BLYP-DFT-D3/DZVP method as implemented in CP2K. While optimizing I have observed that,  SCF energy is decreasing (i.e. becoming more negative) at the beginning of the run, but after few outer-SCF loops, the energy starts increasing and the increment is going on and on if I am leaving the system to run.

Also, I have observed that the inner-SCF energy has reached a convergence (for this system) value  of ~ 10^(-5) after 2 outer-SCF cycles  (where my EPS_SCF is 1.0E-6) and after the beginning of the 3rd outer-SCF tolerance has increased to a value of ~ 10^(-3).

I would like to get your suggestions on any way to avoid this convergence issue. The input file and the output file (just before and after the 2nd outer-SCF) are given below (hopping they may be help you in suggesting).

Thank you all,
Sincerely,
Sharma.

P.S: I have checked in the mailing list but I couldn't exactly found any post on this issue (though there are on BOMD-energy increment etc.)

Input_file:

&FORCE_EVAL

  METHOD Quickstep

  &DFT

    CHARGE = 0
    MULTIPLICITY = 2
    LSD # for odd number of electrons and specifying this keyword itself enforces for spin-polarized calculations

    BASIS_SET_FILE_NAME  BASIS_MOLOPT
    POTENTIAL_FILE_NAME  GTH_POTENTIALS

    &MGRID
      CUTOFF 320
      NGRIDS 5
      REL_CUTOFF 40
    &END MGRID

    &QS
      METHOD GPW
      EPS_DEFAULT 1.0E-10
    &END QS

    &SCF

      SCF_GUESS RESTART
      EPS_SCF 1.0E-6
      MAX_SCF 100

      &OUTER_SCF
        EPS_SCF 1.0E-6
        MAX_SCF 500
      &END

      &OT
        MINIMIZER CG
        PRECONDITIONER FULL_ALL
        ENERGY_GAP 0.001
        STEPSIZE 0.05
      &END

    &END SCF

    &XC

      &XC_FUNCTIONAL BLYP
      &END XC_FUNCTIONAL

      &XC_GRID
        XC_DERIV  SPLINE2
        XC_SMOOTH_RHO NN50
      &END XC_GRID

      &vdW_POTENTIAL

        DISPERSION_FUNCTIONAL PAIR_POTENTIAL # POTENTIAL_TYPE is alias to DISPERSION_FUNCTIONAL
 

        &PAIR_POTENTIAL
          TYPE DFTD3
          PARAMETER_FILE_NAME  dftd3.dat
          REFERENCE_FUNCTIONAL BLYP
        &END PAIR_POTENTIAL

      &END vdW_POTENTIAL

    &END XC

    &POISSON
      POISSON_SOLVER WAVELET
      PERIODIC NONE
    &END POISSON

  &END DFT

  &SUBSYS

    &CELL
      ABC 38.0 38.0 38.0
      ALPHA_BETA_GAMMA  90.0 90.0 90.00
      PERIODIC NONE
    &END CELL

    &TOPOLOGY
      &CENTER_COORDINATES
      &END
    &END

    &COORD
     @INCLUDE gqd_270.xyz
    &END COORD

    &KIND H
      BASIS_SET DZVP-MOLOPT-SR-GTH
      POTENTIAL GTH-BLYP-q1
    &END KIND

    &KIND C
      BASIS_SET DZVP-MOLOPT-SR-GTH
      POTENTIAL GTH-BLYP-q4
    &END KIND

    &KIND N
      BASIS_SET DZVP-MOLOPT-SR-GTH
      POTENTIAL GTH-BLYP-q5
    &END KIND

  &END SUBSYS

&END FORCE_EVAL

&GLOBAL
  PROJECT au20_tetra_on_n-gqd
  RUN_TYPE GEO_OPT
  PRINT_LEVEL MEDIUM
&END GLOBAL

&MOTION
  &GEO_OPT
    OPTIMIZER BFGS
    MAX_ITER   5000
    MAX_FORCE  1.00D-4
    TYPE MINIMIZATION
  &END

&END MOTION

###########################################################################################################################

OUTPUT-FILE:

    98 OT LS       0.54E-01    6.9                    -1982.8376246687

  Trace(PS):                                 1174.9999638412
  Electronic density on regular grids:      -1174.9999638398        0.0000361602
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.0000360489
  Total charge density g-space grids:           0.0000360489

    99 OT CG       0.54E-01   12.3     0.00000845     -1982.8383595833 -7.61E-05

  Trace(PS):                                 1174.9999635922
  Electronic density on regular grids:      -1174.9999635907        0.0000364093
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.0000362980
  Total charge density g-space grids:           0.0000362980

   100 OT LS       0.22E+00    6.9                    -1982.8383813757

  *** SCF run NOT converged ***


  Electronic density on regular grids:      -1174.9999635907        0.0000364093
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.0000362980
  Total charge density g-space grids:           0.0000362980

  Overlap energy of the core charge distribution:               0.00005046108667
  Self energy of the core charge distribution:              -4271.91259980407449
  Core Hamiltonian energy:                                   1351.88408937624945
  Hartree energy:                                            1448.17384258600623
  Exchange-correlation energy:                               -510.06161737575133
  Dispersion energy:                                           -0.92214661917833

  Total energy:                                             -1982.83838137566136

  outer SCF iter =    3 RMS gradient =   0.84E-05 energy =      -1982.8383813757

  ----------------------------------- OT ---------------------------------------

  Allowing for rotations:  F
  Optimizing orbital energies:  F
  Minimizer      : CG                  : conjugate gradient
  Preconditioner : FULL_ALL            : diagonalization, state selective
  Precond_solver : DEFAULT
  Line search    : 2PNT                : 2 energies, one gradient
  stepsize       :    0.05000000
  energy_gap     :    0.00100000

  eps_taylor     :   0.10000E-15
  max_taylor     :             4

  mixed_precision    : F

  ----------------------------------- OT ---------------------------------------

  Step     Update method      Time    Convergence         Total energy    Change
  ------------------------------------------------------------------------------

  Trace(PS):                                 1174.9999628452
  Electronic density on regular grids:      -1174.9999628438        0.0000371562
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.0000370449
  Total charge density g-space grids:           0.0000370449

     1 OT CG       0.50E-01   24.7     0.00073898     -1982.8383578477  1.74E-06

  Trace(PS):                                 1174.6089204778
  Electronic density on regular grids:      -1174.6089204763        0.3910795237
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.3910794124
  Total charge density g-space grids:           0.3910794124

     2 OT LS       0.16E-01    6.8                    -1982.7669389193

  Trace(PS):                                 1174.8781005916
  Electronic density on regular grids:      -1174.8781005901        0.1218994099
  Core density on regular grids:             1174.9999998887       -0.0000001113
  Total charge density on r-space grids:        0.1218992986
  Total charge density g-space grids:           0.1218992986

     3 OT CG       0.16E-01   12.3     0.00075231     -1982.7595972134  7.88E-02

[Matt W]

Hi,

this seems to occur in some systems, sorry no deep physical understanding of the problem - could well be that your geometry is not ideal / nearly metallic electronic structure. A sure sign is that the electron density projected onto the RS grid starts to get (significantly - losing noticable numbers of electrons) worse as the calculation goes on.

FULL_ALL preconditioner is quite aggressive - pragmatically try switching to FULL_SINGLE_INVERSE for an ENERGY calculation until (very nearly) converged, then go back to FULL_ALL.  The speed gains (in SCF cycles) are probably only worth it for long MD runs anyway.  

Matt

[SRKC Sharma Yamijala]

Dear Matt,

Thank you for such a quick reply. I will do my calculation by changing the pre-conditioner.

Just I would like to mention that the above two possibilities which you have mentioned are unlikely for my case because the structure is an already optimized one using other method (PBE) and as it is a quantum dot, it can't be metallic. Any ways, I will do my calculation and let you know whether it is working or not.

Thank you once again,

Sincerely,
Sharma.

********************************************************
Chaitanya Sharma,
Prof. Bala and Prof. Pati's groups,
Chemistry and Physics Materials unit,
JNCASR, BANGLORE,
Lab:: (080-2208) 2581, 2809
https://sites.google.com/site/sharmasrkcyamijala/
*********************************************************

--
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com.
To post to this group, send email to cp...@googlegroups.com.
Visit this group at http://groups.google.com/group/cp2k.
For more options, visit https://groups.google.com/groups/opt_out.

[Florian Schiffmann]

Hi,

this problem has indeed to do with the preconditioner. The statement on the website that a 0.001 value for the energy gap should improve convergence is only partially true. The problem is that it is used as a scaling factor, which for well behaved system (wide band gap usually) is not a problem, but can lead to more random results (problems in inversion due to almost linear dependencies) for small band gap systems and thus mess up your electronic structure (see electronic density on regular grids). Changing the value for the energy gap to 0.1 should thus help as well. As always check as well your plane wave convergence, especially for BLYP larger values might be required.

Flo

[Sharma SRK Chaitanya Yamijala]

Dear Flo, Matt and CP2K users,

I have  changed the preconditioner as well as the energy gap as suggested by Flo and Matt and now the runs are exiting normally. When I just changed the energy gap it has worked for some systems but not for all. Whereas, when I used both of them, then it is working for all the systems (till now what I have considered). Yet, I haven't checked the results only by changing the preconditioner to FULL_SINGLE_INVERSE. Once I get them, I will post.

Many thanks to Flo and Matt,

Sincerely,

Sharma.