"Fastest" AIMD simulation settings/parameters

[ASSIDUO Network]

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 

[ASSIDUO Network]

Just replying to this to bump it higher since I haven't received a response yet.

[Nicholas Winner]

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

[Lenard Carroll]

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.

--
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...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/e7981f13-0b87-4703-9b83-1ff7eadd0361n%40googlegroups.com.

[Marcella Iannuzzi]

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. 

Regards

Marcella

[Lenard Carroll]

What extrapolation method would you recommend for AIMD simulations if I use standard diagonalization and Broyden mixing?

In a lot of the CP2K examples, ASPC is recommended for AIMD simulations. 

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/033b73e4-e60f-4e34-b0cb-30471a0e367bn%40googlegroups.com.

[Jürg Hutter]

Hi

you have to provide enough information in order to give an educated answer.
In your case, if you only changed the thermostat and that leads to a 10 fold
slow down of the calculation, you need to investigate the timing output at the
end of the runs.
Is the timing for the thermostat dominating your calculation? This would be
very special and has to be investigated.
Are there side effects, e.g. number of SCF iterations increasing? Also very unusual
and should be investigated.

You can see: no specific question backed with enough information - no good answers.

regards

Juerg Hutter

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of ASSIDUO Network <lenardc...@gmail.com>
Sent: Sunday, June 12, 2022 9:06 PM
To: cp2k
Subject: [CP2K:17134] Re: "Fastest" AIMD simulation settings/parameters

--
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...@googlegroups.com<mailto:cp2k+uns...@googlegroups.com>.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/833a8c5c-373a-482e-9b6b-04cb0fbfcf36n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/833a8c5c-373a-482e-9b6b-04cb0fbfcf36n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Lenard Carroll]

I have attached my general input file, but also my output file which has the timings.

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...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/ZR0P278MB075933A5AEFB5F1587DD5C6B9FAB9%40ZR0P278MB0759.CHEP278.PROD.OUTLOOK.COM.

[Sam Broderick]

Hi

May I ask what the goal is? It appears to be an ordered Au slab with some disordered Au + O.

I am also struggling with Au, so I am very interested!

Thanks

Sam

[Lenard Carroll]

Investigate what happens to the O atoms and how the surface restructures.

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/bc8bca0f-8bc9-41b3-837a-7bcd2abf4babn%40googlegroups.com.

[Jürg Hutter]

Hi

with timings I was referring to the information printed at the end of a properly terminated job.

It looks like your problem is a poorly converging SCF.
Using Langevin dynamics is a good idea for MD on metals (however, it seems you fix many/most atoms?)
but it has to be combined with OT.
A proper setup is not trivial and needs some experience, see also the literature for protocols.

regards

Juerg Hutter

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of Lenard Carroll <lenardc...@gmail.com>
Sent: Monday, June 13, 2022 11:25 AM
To: cp...@googlegroups.com
Subject: Re: [CP2K:17142] Re: "Fastest" AIMD simulation settings/parameters

I have attached my general input file, but also my output file which has the timings.

On Mon, Jun 13, 2022 at 11:13 AM Jürg Hutter <hut...@chem.uzh.ch<mailto:hut...@chem.uzh.ch>> wrote:
Hi

you have to provide enough information in order to give an educated answer.
In your case, if you only changed the thermostat and that leads to a 10 fold
slow down of the calculation, you need to investigate the timing output at the
end of the runs.
Is the timing for the thermostat dominating your calculation? This would be
very special and has to be investigated.
Are there side effects, e.g. number of SCF iterations increasing? Also very unusual
and should be investigated.

You can see: no specific question backed with enough information - no good answers.

regards

Juerg Hutter

________________________________________

From: cp...@googlegroups.com<mailto:cp...@googlegroups.com> <cp...@googlegroups.com<mailto:cp...@googlegroups.com>> on behalf of ASSIDUO Network <lenardc...@gmail.com<mailto:lenardc...@gmail.com>>

Sent: Sunday, June 12, 2022 9:06 PM
To: cp2k
Subject: [CP2K:17134] Re: "Fastest" AIMD simulation settings/parameters

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 the Google Groups "cp2k" group.

To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+uns...@googlegroups.com<mailto:cp2k%2Bunsu...@googlegroups.com><mailto:cp2k+uns...@googlegroups.com<mailto:cp2k%2Bunsu...@googlegroups.com>>.

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/833a8c5c-373a-482e-9b6b-04cb0fbfcf36n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/833a8c5c-373a-482e-9b6b-04cb0fbfcf36n%40googlegroups.com?utm_medium=email&utm_source=footer>.

--
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...@googlegroups.com<mailto:cp2k%2Bunsu...@googlegroups.com>.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/ZR0P278MB075933A5AEFB5F1587DD5C6B9FAB9%40ZR0P278MB0759.CHEP278.PROD.OUTLOOK.COM.

--
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...@googlegroups.com<mailto:cp2k+uns...@googlegroups.com>.

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/CAKur15-8pk3RVx_e%3D%3DLXRwFZsfmty1N5v1hT%3Dt__aXoZ8HSZAg%40mail.gmail.com<https://groups.google.com/d/msgid/cp2k/CAKur15-8pk3RVx_e%3D%3DLXRwFZsfmty1N5v1hT%3Dt__aXoZ8HSZAg%40mail.gmail.com?utm_medium=email&utm_source=footer>.