Setting correct cutoff/convergence criteria

[ashish dabral]

Hi,

I am using cp2k for geometry optimization (GEO_OPT only). I have a system with almost 300 atoms(40x25x11 CELL). Even though cp2k is really fast, it seems

like here it is taking too much time as I have to wait almost 2 weeks to get it to converge with 12 cores (it does converge!) and a gamma point calculation.

I am wondering if my cut-off/convergence values are ok. I realized that the GEO_OPT parameter section was missing in the MOTION section but I assume

then that the default values were used. There is a variety of atoms in the complex (Si, C, Mo, Br etc.). Is it normal then, given the below setup, to consume 

so much time? A complete input is not possible to attach but I am attaching the file showing different parameters: 

&FORCE_EVAL

  METHOD QS

   &DFT

    MULTIPLICITY  1

    CHARGE  0

    BASIS_SET_FILE_NAME /BASIS_MOLOPT

    POTENTIAL_FILE_NAME /GTH_POTENTIALS

    SURFACE_DIPOLE_CORRECTION TRUE

    SURF_DIP_DIR Z

    &QS

      METHOD GPW

      EXTRAPOLATION USE_GUESS

    &END QS

    &MGRID

      CUTOFF 900

      REL_CUTOFF 60

    &END MGRID

    &PRINT

      &MO OFF

      &END MO

       &PDOS DEBUG 

          NLUMO 5

       &END PDOS

    &END PRINT

    &SCF

      SCF_GUESS ATOMIC

      MAX_SCF 400

      ADDED_MOS 200

      &DIAGONALIZATION T

        ALGORITHM STANDARD

      &END DIAGONALIZATION

      &MIXING T

        NBUFFER 4

        BETA 1.25

        ALPHA 0.2

        METHOD BROYDEN_MIXING

      &END MIXING

      &PRINT

        &RESTART ON

        &END RESTART

      &END PRINT

      &SMEAR ON

        METHOD FERMI_DIRAC

        ELECTRONIC_TEMPERATURE [K] 300

      &END SMEAR

    &END SCF

    &XC

      &XC_FUNCTIONAL PBE

        &PBE

          PARAMETRIZATION ORIG

        &END PBE

      &END XC_FUNCTIONAL

      &VDW_POTENTIAL                                                                                     

        POTENTIAL_TYPE PAIR_POTENTIAL                                                                       

        &PAIR_POTENTIAL                                                                                                    

            TYPE DFTD3                                                                                  

            REFERENCE_FUNCTIONAL PBE                                                                       

            R_CUTOFF  10.0                                                                                   

            PARAMETER_FILE_NAME /dftd3.dat                       

            SCALING 0.0                                                                                         

        &END PAIR_POTENTIAL                                                                                      

      &END VDW_POTENTIAL 

    &END XC

  &END DFT

  &PRINT

      &STRESS_TENSOR ON

      &END STRESS_TENSOR

      &FORCES ON

      &END FORCES

  &END PRINT 

&END FORCE_EVAL

&MOTION

    &CELL_OPT

      OPTIMIZER BFGS

      MAX_DR 0.005

      MAX_ITER 100

      EXTERNAL_PRESSURE 0

      MAX_FORCE 0.0004

      RMS_DR 0.002

      PRESSURE_TOLERANCE 100

      RMS_FORCE 0.001

      TYPE DIRECT_CELL_OPT

    &END CELL_OPT                  

  &PRINT 

      &TRAJECTORY

          FORMAT PDB

      &END TRAJECTORY

  &END PRINT

&END MOTION               

Thanks for any inputs.

[Matt W]

Hi,

there are quite a few things that could be improved in your input, but the most significant problem is probably

   EXTRAPOLATION USE_GUESS

This means that after every step it starts again from an atomic guess, which is very inefficient. By default CP2K extrapolates from the density matrix of the previous steps.

Also, 900 Ry is a very expensive cutoff (might be needed for some elements with GPW). Change to METHOD GAPW in &QS and reduce the cutoff to ~300 Ry.

Mixing and broyden optimizer will be moderately inefficient for insulators, but the above are probably the biggest things.

Hope that helps,

Matt

[ashish dabral]

Hi Matt,

Thanks for your suggestions. After playing around with the parameters, I figured the Extrapolation issue and cut-off issue (so fixed that). 

I already see improvement in time after some trials. I am wondering though which method would be best for extrapolation: ASPC etc.

I haven't switched to GAPW though. I will try that. Can you please elaborate on "Mixing and broyden optimizer will be moderately inefficient for insulators"

and a possible alternative.

Thanks

[Matt W]

Thanks for your suggestions. After playing around with the parameters, I figured the Extrapolation issue and cut-off issue (so fixed that). 

I already see improvement in time after some trials. I am wondering though which method would be best for extrapolation: ASPC etc.

I'd stick with ASPC. PS is maybe a bit faster for geo_opt, but ASPC should be good and robust.

 

I haven't switched to GAPW though. I will try that.

Only needed really when you have some elements that need stupidly large cutoff (F, Na, Mg) with normal pseudos. It does tend to make forces a bit more accurate, so can speed up final convergence a lot too.

 

Can you please elaborate on "Mixing and broyden optimizer will be moderately inefficient for insulators"

and a possible alternative.

The Orbital Transform (OT) method is CP2Ks flagship optimizer for insulators. For a smallish system

    &SCF

      &OUTER_SCF

         EPS_SCF 5.0E-7

         MAX_SCF 1

      &END OUTER_SCF

      SCF_GUESS RESTART

      EPS_SCF 5.0E-7

      MAX_SCF  15

      &OT

        MINIMIZER DIIS

        PRECONDITIONER FULL_ALL

        ENERGY_GAP 0.001

      &END

    &END SCF

would be my default setup. 

        PRECONDITIONER FULL_SINGLE_INVERSE

        ENERGY_GAP 0.1

for large systems (>500 atoms or so)

Matt

 

Thanks

[ashish dabral]

Thanks a lot Matt for your suggestions and additional clarification.