[CP2K-user] need suggestions to speed up hybrid functionals with ADMM

[Geng Sun]

Hello Matt,

Thank you very much.

The EPS_SCHWARZ indeed has a significant impact on the CPU time for an SCF 
loop.
I decreased EPS_SCHWARZ and it is much faster.

Now I did not notice any convergence problem in the test system.

Best Wishes,

Geng

在 2019年9月23日星期一 UTC-7上午10:01:38，Matt W写道：
>
> Hi,
>
> to get reasonable hybrid performance you need to get all the two-electron 
> integrals into memory. Then the first scf step will be much longer than the 
> others.
>
> To try and reduce the number of two-electron integrals you need to screen 
> out the many tiny ones.
>
>             &SCREENING
>                EPS_SCHWARZ 1.0E-12
>                SCREEN_ON_INITIAL_P FALSE
>             &END SCREENING
>
> try changing to 
>
>             &SCREENING
>                EPS_SCHWARZ 1.0E-6
>                SCREEN_ON_INITIAL_P TRUE
>             &END SCREENING
>
> and restart from a converged GGA calculation. If it doesn't converge you 
> may need to decrease EPS_SCHWARZ
>
> EPS_SCHWARZ should really be a convergence parameter, but if the SCF 
> works, probably it is small enough. The Aluminium ADMM basis might be a 
> problem too, but see if this helps.
>
> Matt
>
> On Sunday, September 22, 2019 at 2:23:23 AM UTC+1, Geng Sun wrote:
>>
>> Dear CP2K users,
>>
>> I am trying to optimize alumina supported Cu oxide clusters with hybrid 
>> functional (hse06), 
>> The ADMM method looks very promising because of its efficiency.
>>
>> The system I am studying is rather large, which contains about 240 atoms 
>> for substrate(Al2O3)  and another 10 atoms for CuxOy clusters.
>> Below is the input I used for the calculation.
>>
>> Now the difficulty is that a single SCF step (an electronic step) with 
>> this input still needs about  2~3 hours (the first one takes 7 hours).  (I 
>> am using 256 CPUs for this calculation). Hence it is very difficult to 
>> carry out further optimizations with this setup. 
>>
>> As a reference, I also used VASP to conduct HSE06 calculation for this 
>> structure, and VASP takes 1 hour to complete an SCF step and gets 
>> convergence in 30 steps (with also 256 CPUs)  
>>
>> My question is  whether my test is reliable or not? If it is not 
>> consistent with the general impression of experienced users, what can be 
>> improved? 
>>
>> Thank you very much in advance.
>>
>> Best,
>>
>> Geng
>>
>>
>> &FORCE_EVAL
>>    &DFT
>>       UKS
>>       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
>>       &MGRID
>>          CUTOFF 400
>>          COMMENSURATE
>>       &END MGRID
>>       &QS
>>          EXTRAPOLATION PS
>>          EXTRAPOLATION_ORDER 3
>>          EPS_DEFAULT  1.0E-12
>>          MAP_CONSISTENT .True.
>>          EPS_PGF_ORB 1.0E-32
>>       &END QS
>>       &AUXILIARY_DENSITY_MATRIX_METHOD
>>          METHOD   BASIS_PROJECTION
>>          ADMM_PURIFICATION_METHOD MO_DIAG
>>       &END AUXILIARY_DENSITY_MATRIX_METHOD
>>       &SCF
>>          EPS_SCF 5.0E-7
>>          MAX_SCF 30
>>          SCF_GUESS RESTART
>>          &OT
>>            ENERGY_GAP 0.002
>>            LINESEARCH 3PNT
>>            PRECONDITIONER FULL_ALL
>>            MINIMIZER CG
>>          &END OT
>>          &OUTER_SCF
>>            EPS_SCF 5.0E-7
>>            MAX_SCF 40
>>          &END OUTER_SCF
>>       &END SCF
>>       &XC
>>          !&XC_GRID
>>          !    USE_FINER_GRID T
>>          !&END XC_GRID
>>          &XC_FUNCTIONAL
>>             &PBE
>>                SCALE_X 0.0
>>                SCALE_C 1.0
>>             &END PBE
>>             &XWPBE
>>                SCALE_X -0.25
>>                SCALE_X0 1.0
>>                OMEGA 0.11
>>             &END XWPBE
>>          &END XC_FUNCTIONAL
>>          &HF
>>             FRACTION 0.25
>>             &SCREENING
>>                EPS_SCHWARZ 1.0E-12
>>                SCREEN_ON_INITIAL_P FALSE
>>             &END SCREENING
>>             &INTERACTION_POTENTIAL
>>                CUTOFF_RADIUS 6.0
>>                POTENTIAL_TYPE SHORTRANGE
>>                OMEGA 0.11
>>             &END INTERACTION_POTENTIAL
>>             &MEMORY
>>                MAX_MEMORY  2400
>>                EPS_STORAGE_SCALING 0.1
>>             &END MEMORY
>>          &END HF
>>       &END XC
>>       &PRINT
>>          &MO
>>              OCCUPATION_NUMBERS TRUE
>>              &EACH
>>               QS_SCF  0
>>              &END EACH
>>         &END MO
>>       &END PRINT
>>    &END DFT
>>    &SUBSYS
>>       &KIND Cu
>>          BASIS_SET TZV2P-MOLOPT-SR-GTH
>>          BASIS_SET AUX_FIT FIT11
>>          POTENTIAL GTH-PBE-q11
>>       &END KIND
>>       &KIND O
>>          BASIS_SET TZVP-MOLOPT-GTH
>>          BASIS_SET AUX_FIT pFIT3
>>          POTENTIAL GTH-PBE-q6
>>       &END KIND
>>       &KIND Al
>>          BASIS_SET TZVP-MOLOPT-SR-GTH
>>          BASIS_SET AUX_FIT FIT9
>>          POTENTIAL GTH-PBE-q3
>>       &END KIND
>>    &END SUBSYS
>>    &PRINT
>>       &FORCES ON
>>          LOG_PRINT_KEY True
>>       &END FORCES
>>    &END PRINT
>> &END FORCE_EVAL
>>
>>
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20190924/f715a661/attachment.htm>