[CP2K-user] [CP2K:16266] Re: Improve extrapolation between relaxation steps?
Marcella Iannuzzi
marci.akira at gmail.com
Thu Nov 18 09:17:55 UTC 2021
Dear Nick
Being possibly a spin polarised, charged, and small gap system it is not
that surprising that the SCF has troubles.
Does the total energy fluctuates significantly among the optimisation
steps?
Have you tried to bias the magnetisation of the atoms close to the defect
to help the SCF find solutions for the real space distribution?
It happens that the SCF gets trapped into a wrong minimum and then
convergence becomes unstable. Have you checked whether the charge
distribution you get at any of those SCF steps is what you would expect?
Best
Marcella
On Wednesday, November 17, 2021 at 7:25:06 PM UTC+1 nwi... at berkeley.edu
wrote:
>
> It is a large band gap (TiO2), but with a has a shallow defect state,
> making the overall band gap small (~0.13eV).
>
> I know OT works better the wider the gap, but I had hooped that after the
> first SCF, it would still speed up quite a bit for the subsequent
> relaxation steps.
>
> I was trying to use FULL_KINETIC because
> IRAC+Rotations+occupation_preeconditioner together require that. This has
> the advantage of making defect states more robust by allowing for
> fractional occupations, but of course it can severely compromise
> convergence speed.
>
> When I switch to FULL_SINGLE_INVERSE preconditioner. I can grep and get:
>
> * outer SCF loop converged in* 3 iterations or 104 steps
>
> *outer SCF loop converged in* 2 iterations or 72 steps
>
> *outer SCF loop converged in* 3 iterations or 116 steps
>
> *outer SCF loop converged in* 2 iterations or 72 steps
>
> *outer SCF loop converged in* 2 iterations or 93 steps
>
> *outer SCF loop converged in* 2 iterations or 72 steps
>
> *outer SCF loop converged in* 2 iterations or 63 steps
>
> *outer SCF loop converged in* 2 iterations or 60 steps
>
>
> It doesn't look like its speeding up very much. I suppose 60 i better than
> 100, but its not what I expected.
> If this is just the way it is, then I'll live with it.
>
> On Wednesday, November 17, 2021 at 10:08:02 AM UTC-8 mattwa... at gmail.com
> wrote:
>
>> 'so for wide gap insulators you should for sure only need a few SCF
>> cycles to converge'
>> I meant to write when fairly relaxed / close to the minimum ...
>>
>>
>>
>> On Wednesday, 17 November 2021 at 18:06:33 UTC Matt Watkins wrote:
>>
>>> Well, CP2K was setup to do Born-Oppenheimer MD for bio systems - so
>>> for wide gap insulators you should for sure only need a few SCF cycles to
>>> converge.
>>>
>>> Is your system insulating (band gap, let's say > 0.2 eV)? If not
>>> different approach needed.
>>>
>>> If so, improve the OT setup - especially the preconditioner
>>>
>>> remove
>>> PRECONDITIONER FULL_KINETIC
>>> ENERGY_GAP 0.01
>>>
>>> try
>>>
>>> PRECONDITIONER FULL_SINGLE_INVERSE
>>>
>>> or (can be expensive for larger systems)
>>>
>>> PRECONDITIONER FULL_ALL
>>>
>>> and leave out the energy_gap (if you are using an up-to-date cp2k)
>>>
>>> Also check that you XC setup is correct ...
>>>
>>> Matt
>>>
>>>
>>> On Tuesday, 16 November 2021 at 17:58:59 UTC nwi... at berkeley.edu wrote:
>>>
>>>> Hello all,
>>>>
>>>> When I run a relaxation for a complex system, e.g. in my case a
>>>> defective transition metal oxide with hybrid dft on a modest ~200 atom
>>>> supercell, each SCF still takes quite a long time to converge.
>>>>
>>>> I originally came from vasp before switching to cp2k, and it in that
>>>> code, my experience was that once the first scf loop converged, the next
>>>> relaxation step would take much less time, and so on until the last few
>>>> steps converged in just a few iterations. This isn't my experience with
>>>> CP2K though. Running CG on my most recent relaxation I still find that
>>>> after more than 20 relaxation steps, 40-50 scf steps are required to
>>>> achieve convergence.
>>>>
>>>> Is this something that I could improve by changing certain settings?
>>>> Below is the relevant info for my input file. Please note that I understand
>>>> FULL_KINETIC and CG might lead to slower convergence than FULL_ALL + DIIS,
>>>> but my question lies in why convergence is not appreciably speeding up
>>>> rather than its absolute speed.
>>>>
>>>> -Nick
>>>>
>>>> &DFT
>>>>
>>>> BASIS_SET_FILE_NAME BASIS_MOLOPT
>>>>
>>>> BASIS_SET_FILE_NAME BASIS_MOLOPT_UCL
>>>>
>>>> BASIS_SET_FILE_NAME BASIS_ADMM
>>>>
>>>> BASIS_SET_FILE_NAME BASIS_ADMM_MOLOPT
>>>>
>>>> POTENTIAL_FILE_NAME GTH_POTENTIALS
>>>>
>>>> UKS T
>>>>
>>>> MULTIPLICITY 0
>>>>
>>>> CHARGE 4
>>>>
>>>> &SCF
>>>>
>>>> MAX_ITER_LUMO 400
>>>>
>>>> MAX_SCF 50
>>>>
>>>> EPS_SCF 1e-06
>>>>
>>>> SCF_GUESS RESTART
>>>>
>>>> &OT
>>>>
>>>> ALGORITHM STRICT
>>>>
>>>> MINIMIZER CG
>>>>
>>>> LINESEARCH 2PNT
>>>>
>>>> PRECONDITIONER FULL_KINETIC
>>>>
>>>> ENERGY_GAP 0.01
>>>>
>>>> ROTATION F
>>>>
>>>> OCCUPATION_PRECONDITIONER F
>>>>
>>>> &END OT
>>>>
>>>> &OUTER_SCF
>>>>
>>>> EPS_SCF 1e-06
>>>>
>>>> MAX_SCF 20
>>>>
>>>> &END OUTER_SCF
>>>>
>>>> &END SCF
>>>>
>>>> &AUXILIARY_DENSITY_MATRIX_METHOD
>>>>
>>>> ADMM_PURIFICATION_METHOD None
>>>>
>>>> METHOD BASIS_PROJECTION
>>>>
>>>> &END AUXILIARY_DENSITY_MATRIX_METHOD
>>>>
>>>> &QS
>>>>
>>>> EPS_DEFAULT 1e-12
>>>>
>>>> EPS_PGF_ORB 1e-16
>>>>
>>>> EXTRAPOLATION PS
>>>>
>>>> METHOD GPW
>>>>
>>>> &END QS
>>>>
>>>> &MGRID
>>>>
>>>> NGRIDS 5
>>>>
>>>> CUTOFF 550.0
>>>>
>>>> REL_CUTOFF 50.0
>>>>
>>>> &END MGRID
>>>>
>>>> &XC
>>>>
>>>> &XC_FUNCTIONAL NO_SHORTCUT
>>>>
>>>> &PBE
>>>>
>>>> SCALE_X 0.0
>>>>
>>>> SCALE_C 1.0
>>>>
>>>> &END PBE
>>>>
>>>> &XWPBE
>>>>
>>>> SCALE_X 0.063
>>>>
>>>> SCALE_X0 0.877
>>>>
>>>> OMEGA 0.07
>>>>
>>>> &END XWPBE
>>>>
>>>> &PBE_HOLE_T_C_LR
>>>>
>>>> SCALE_X 0.123
>>>>
>>>> CUTOFF_RADIUS 5.930000000000001
>>>>
>>>> &END PBE_HOLE_T_C_LR
>>>>
>>>> &END XC_FUNCTIONAL
>>>>
>>>> &HF
>>>>
>>>> FRACTION 1.0
>>>>
>>>> &SCREENING
>>>>
>>>> EPS_SCHWARZ 1e-07
>>>>
>>>> EPS_SCHWARZ_FORCES 1e-06
>>>>
>>>> SCREEN_P_FORCES T
>>>>
>>>> SCREEN_ON_INITIAL_P T
>>>>
>>>> &END SCREENING
>>>>
>>>> &INTERACTION_POTENTIAL
>>>>
>>>> POTENTIAL_TYPE MIX_CL_TRUNC
>>>>
>>>> OMEGA 0.07
>>>>
>>>> SCALE_COULOMB 0.06
>>>>
>>>> SCALE_LONGRANGE 0.063
>>>>
>>>> CUTOFF_RADIUS 5.930000000000001
>>>>
>>>> T_C_G_DATA t_c_g.dat
>>>>
>>>> &END INTERACTION_POTENTIAL
>>>>
>>>> &LOAD_BALANCE
>>>>
>>>> RANDOMIZE T
>>>>
>>>> &END LOAD_BALANCE
>>>>
>>>> &MEMORY
>>>>
>>>> EPS_STORAGE_SCALING 0.1
>>>>
>>>> MAX_MEMORY 2000
>>>>
>>>> &END MEMORY
>>>>
>>>> &END HF
>>>>
>>>> &END XC
>>>>
>>>> &END DFT
>>>>
>>>
--
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+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/9ba22811-b838-48e5-aac1-4d89b91d146dn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20211118/48b29a10/attachment.htm>
More information about the CP2K-user
mailing list