[CP2K-user] [CP2K:17140] Re: "Fastest" AIMD simulation settings/parameters

Marcella Iannuzzi marci.akira at gmail.com
Mon Jun 13 08:29:14 UTC 2022

If you are using Broyden mixing, the optimisation of the wfn is most 
probably based on standard diagonalization, hence no preconditioner is 
applied and neither CG nor DIIS as well. 
With standard diagonalization+Broyden mixing ASPC is also not really the 
best choice. 

On Monday, June 13, 2022 at 10:17:59 AM UTC+2 lenardc... at gmail.com wrote:

> It seems I am already using pretty much the best (fastest) settings 
> possible then.
> For my AIMD simulations I do use Broyden Mixing, with a large-ish alpha 
> value, FULL_KINETIC predconditioner, ASPC extrapolation method, no UKS and 
> a 2 fs time step. I am, however, using CG for the minimizer instead of 
> DIIS. I am going to test that out and see how it looks.
> Thank you very much Nick.
> On Mon, Jun 13, 2022 at 12:48 AM Nicholas Winner <nwi... at berkeley.edu> 
> wrote:
>> Here's a list of some things I can think of to increase speed. Probably 
>> not comprehensive. In the future include your input file for people so we 
>> can see what settings you are already trying.
>> Will not compromise accuracy (test all options to find best)
>>    1. Using traditional diagonalization:
>>       1. Use broyden mixer, which is generally (not always!) faster. 
>>       2. In broyden use as large of a alpha value as you can afford 
>>       while still achieving convergence.
>>    2. Using OT:
>>       1. Change preconditioner to speed up the scf convergence. Normally 
>>       Full_all or Full_single_inverse are most effective, but full_kinetic might 
>>       be good if you have a huge system that takes too long to precondition.
>>       2. if you know the system energy gap, you can supply a slight 
>>       underestimate of it to the OT preconditioner to help it converge faster
>>       3. Change from CG (which is more robust) to DIIS which is usually 
>>       faster
>>    3. Use ASPC wfn extrapolation
>>    4. Inspect your SCF convergence. If you see unstable long 
>>    convergence, then increasing your precision settings might help you in the 
>>    long run.
>> Will/might compromise accuracy
>>    1. Decrease EPS_DEFAULT 
>>    2. Decrease EPS_SCF (tolerance for scf convergence)
>>    3. Use UKS = False. I.e. spin unpolarized if you not need it.
>>    4. Use a larger MD time-step. Some systems are still okay with a 
>>    large time step of 2fs, could work for you.
>> -Nick
>> On Sunday, June 12, 2022 at 12:06:49 PM UTC-7 lenardc... at gmail.com wrote:
>>> Just replying to this to bump it higher since I haven't received a 
>>> response yet.
>>> On Thursday, June 9, 2022 at 3:38:02 AM UTC+2 ASSIDUO Network wrote:
>>>> Hi there, I have been running AIMD simulations for about a year now and 
>>>> it originally started off really slow, only getting 60 fs per day. In 
>>>> particular, I used the NVT ensemble and the Nose-Hoover thermostat. When 
>>>> swapping to the Langevin ensemble and adaptive-Langevin thermostat, the 
>>>> simulation time improved to 200-600 fs a day. 
>>>> I wanted to know what settings/parameters you would suggest to improve 
>>>> the speed even further? That is, what are the best settings to choose to 
>>>> get the fastest performance?
>>>> Key things to keep in mind for my work:
>>>> Using GTH-PBE potential and the DZVP-MOLOPT-SR-GTH basis set. Fermi 
>>>> dirac smearing must also be used 
>>> -- 
>> You received this message because you are subscribed to a topic in the 
>> Google Groups "cp2k" group.
>> To unsubscribe from this topic, visit 
>> https://groups.google.com/d/topic/cp2k/I0mGOcOFu6s/unsubscribe.
>> To unsubscribe from this group and all its topics, send an email to 
>> cp2k+uns... at googlegroups.com.
>> To view this discussion on the web visit 
>> https://groups.google.com/d/msgid/cp2k/e7981f13-0b87-4703-9b83-1ff7eadd0361n%40googlegroups.com 
>> <https://groups.google.com/d/msgid/cp2k/e7981f13-0b87-4703-9b83-1ff7eadd0361n%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/033b73e4-e60f-4e34-b0cb-30471a0e367bn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220613/f3d7ff7e/attachment-0001.htm>

More information about the CP2K-user mailing list