[CP2K-user] [CP2K:18938] Increase of the SCF cycles along an AIMD run

[Dune André]

Hi Lucas,

Thanks for your answer. I have tested to increase the SCF convergence 
thresholds as you mentioned it: one calculation with (EPS_DEFAULT=10-10 et 
EPS_SCF=10-5) and one calculation with (EPS_DEFAULT=10-12 et EPS_SCF=10-6). 
In the former calculation, I had the same problem (increase of the SCF 
cycles along the trajectory) while in the latter SCF did not converge. 

I tried improving SCF convergence by using an outer SCF loop but it was 
unsuccessful. The trajectory does not look unphysical, I don't see weird 
structures.

It seems that the increasing number of SCF cycles comes from the 
combination of DIIS method and FULL_KINETIC preconditioner (maybe due to a 
numerical effect?) but I am stuck to overcome this difficulty. Do you see 
other potential reasons of this behavior? 

Thanks a lot.

Best regards,

Dune


Le lundi 5 juin 2023 à 22:22:38 UTC+2, Lucas Lodeiro a écrit :

> As a first glance, your EPS_DEFAULT is 10⁻⁷ which is too big, I usually 
> use 10⁻¹⁰ at least, or 10⁻¹² for more precise AIMD. Also a good relation 
> between EPS_DEFAULT and EPS_SCF is to set the last one as the square root 
> of the former.
> All this affects the quality of the forces which dominates the MD... maybe 
> if you check the geometry of the long SCF frames you could find some weird 
> bond lengths or unphysical structures which can explain the behaviour.
>
> Regards - Lucas
>
> El lun, 5 jun 2023 a las 14:09, Dune André (<dune.... at gmail.com>) 
> escribió:
>
>> Dear all,
>>
>> I am facing an issue while running AIMD simulations of a box of ~1000 
>> atoms. The system is an amorphous MOF made of Fe linked by benzene 
>> tricarboxylate (named Fe-BTC).
>>
>> I use PBE functional with DZVP basis set and GTH pseudopotential ; DIIS 
>> method and FULL_KINETIC preconditioner. I perform the calculation from a 
>> restart file. The first steps of the MD are quite fast with ~10 SCF steps 
>> per MD step. But along the AIMD run, the SCF cycles become longer and 
>> longer (~100 SCF steps per MD step). 
>>
>> I tried to use other OT method such as CG or other preconditioner such as 
>> FULL_SINGLE_INVERSE but in both cases the SCF cycles are taking too much 
>> time and sometimes fail to converge. Besides, FULL_ALL preconditioner is 
>> too computationally expensive in terms of memory.
>>
>> Do you have some ideas of the origin of this increase of the SCF cycles? 
>> I attached the restart file to the message. 
>>
>> Thanks in advance for your help,
>>
>> Kind regards,
>>
>> Dune André
>>
>> -- 
>> 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/5d10d338-5a72-475a-adb4-d4ea7ecbd2d2n%40googlegroups.com 
>> <https://groups.google.com/d/msgid/cp2k/5d10d338-5a72-475a-adb4-d4ea7ecbd2d2n%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/38e99e67-10bc-45d6-ac22-7bb4a550ca80n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20230619/2e109f10/attachment.htm>