[CP2K-user] [CP2K:18251] Re: How to speed HF up

[DMT]

Thank you for the suggestions.

Yes, when I try with a PBE GUESS the convergence is a lot faster. It starts 
with a lower Energy and improves by all factors much faster with the 
micro-iterations.

In my experience so far diagonalization is faster than OT. I usually try 
both.

>From my testing Smearing does help with metal systems like graphene, but 
for organic molecules it only slows things down.

>From my testing Mixing can help with larger systems, but isn't Mixing only 
for periodic conditions and not supposed to be used with gas phase organics 
?

Regards,
Dobromir

On Thursday, December 22, 2022 at 10:28:30 PM UTC+2 nwi... at berkeley.edu 
wrote:

> 1. EPS_DEFAULT should be increased to at least 1e-10, maybe 1e-12. It does 
> not change the speed of the HF much when you reduce it and creates 
> substantial errors.
> 2. Use a PBE restart file and then set SCREEN_ON_INITIAL_P True. This 
> creates very effective screening and rarely causes problems. The initial 
> wave function guess needs to be "good" for this approximation to be valid, 
> but in my experience it is only invalid when one is very far from 
> equilibrium geometries.
> 3. For your auxiliary basis set, admm-dzp should already be fairly fast, 
> but if you want to be more aggressive you can use the older cpFIT basis 
> sets. These are less diffuse and will result in quicker HF evaluation.
>
> Other than that there aren't many ways you can squeeze performance other 
> than changing your SCF section. Are you noticing that you are taking many 
> many SCF steps to converge? Maybe your SCF settings are the problem. I see 
> that you are using diagonalization instead of OT. If you must use 
> traditional diagonalization, that is okay, but I would suggest trying to 
> use some mixing scheme. The default mixing is good for some systems, but in 
> general I find broyden mixing is better. Something like
>
> &SCF
>             MAX_SCF 100                    !DEFAULT 50
>             EPS_SCF 1.0E-5                !DEFAULT 1.0e-5, OT: -6 - -7, 
> FOR DIAG CAN BE SMALLER (DEFAULT)
>             SCF_GUESS RESTART            !DEFAULT ATOMIC, RESTART FOR 
> RESTART
>             &DIAGONALIZATION T
>             &END DIAGONALIZATION
>             &MIXING
>                     METHOD BROYDEN_MIXING
>                     ALPHA .01
>                     BETA .1
>              &END
> &END SCF
>
>
> If you include your output it can also help us diagnose what is the 
> problem and if there are improvements. 
>
> A side note to Augustine's suggestion. RI will not benefit ADMM. RI can be 
> used to improve the speed of HFX with large basis, but since ADMM is a 
> small basis the improvements go away.
>
> On Thursday, December 22, 2022 at 4:34:02 AM UTC-8 DMT wrote:
>
>> Hi Augustin,
>>
>> Thank you for the conformation and advises. I will try RI
>>
>> Best Regards,
>> Dobromir
>>
>> On Thursday, December 22, 2022 at 1:30:57 PM UTC+2 Augustin Bussy wrote:
>>
>>> Hi Dobromir,
>>>
>>> you should keep in mind that HFX is inherently an expensive method 
>>> (compared to GGA). From your input file, I would say you are already using 
>>> all the tricks I know of:
>>>
>>>    1. Using the ADMM  method
>>>    2. Using a truncated Coulomb operator, although I have to warn you 
>>>    about using extreme values such as 2.0 Angstrom (~6 is usually good)
>>>    3. Using the largest possible value for MAX_MEMORY
>>>    4. Using a loose EPS_SCHWARZ threshold. Here too, 1.0E-6, may be on 
>>>    the limit
>>>
>>> Also, I think that your EPS_DEFAULT value might be dangerously loose, if 
>>> you want accurate results. You could optionally try to use the RI-HFX 
>>> method by adding the &RI subsection in &HF. Default parameters should be 
>>> good enough. There is however no guarantee that it speeds up you particular 
>>> calculation.
>>>
>>> Best,
>>> Augustin
>>>
>>>
>>> ------------------------------
>>> *From:* cp... at googlegroups.com <cp... at googlegroups.com> on behalf of 
>>> DMT <dobro... at gmail.com>
>>> *Sent:* Thursday, December 22, 2022 12:02 PM
>>> *To:* cp2k <cp... at googlegroups.com>
>>> *Subject:* [CP2K:18243] Re: How to speed HF up 
>>>  
>>> I tried  
>>>
>>>                 !&PBE_HOLE_T_C_LR
>>>                 !    CUTOFF_RADIUS 6.0
>>>                 !    SCALE_X 0.25
>>>                 !&END PBE_HOLE_T_C_LR
>>>
>>> in &XF_FUNCTIONAL and
>>>
>>>             &HF
>>>                 FRACTION 0.25
>>>                 &SCREENING
>>>                     EPS_SCHWARZ 1.0E-6            !DEFAULT 1.0E-10, 
>>> TYPICAL -6 - -9, FOR STABLE HFX
>>>                     !SCREEN_ON_INITIAL_P T         !FOR A GOOD INITIAL 
>>> GUESS FROM A .WFN
>>>                 &END SCREENING
>>>                 &INTERACTION_POTENTIAL
>>>                     POTENTIAL_TYPE TRUNCATED    !TRUNCATED 1/r (Coulomb) 
>>> POTENTIAL FOR SPEED UP
>>>                     CUTOFF_RADIUS 6.0            !SHOULD BE < HALF THE 
>>> CELL
>>>                     T_C_G_DATA /home/lsd/cp2k_data/t_c_g.dat     !DATA 
>>> FOR THE EVAL OF TRUNCATED GAMMA FUNC
>>>                 &END INTERACTION_POTENTIAL
>>>
>>> in &XC and there is no significant (if any) speed up.
>>>
>>> I even played with the CUTOFF_RADIUS (values from 2.0 to 6.0) and 
>>> nothing changes.
>>>
>>> Am I missing something ?
>>>
>>> Best Regards,
>>> Dobromir
>>>
>>> On Tuesday, December 20, 2022 at 8:31:44 PM UTC+2 DMT wrote:
>>>
>>> Hello, 
>>>
>>> I have heard that there must be a way to speed HF (for hybrid DFT 
>>> methods) up. Can you help me with that ?
>>>
>>> Best Regards,
>>> Dobromir
>>>
>>> -- 
>>> 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... at googlegroups.com.
>>> To view this discussion on the web visit 
>>> https://groups.google.com/d/msgid/cp2k/64f47d87-163a-4a6d-8061-c9b0b660b928n%40googlegroups.com 
>>> <https://groups.google.com/d/msgid/cp2k/64f47d87-163a-4a6d-8061-c9b0b660b928n%40googlegroups.com?utm_medium=email&utm_source=footer>
>>> .
>>>
>>

-- 
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/a120c669-bf60-4a3f-9190-2b4aeccc5d00n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20221224/fc4aa0bb/attachment.htm>