Convergence issue for GAPW without ADMM

[ma455...@gmail.com]

Hi,

I'm a new user of cp2k 8.1. I'm currently using ALL electron basis set and GAPW method to do a simple Si supercell calculation. But without ADMM, the SCF process does not converge. I do not want to use ADMM in this case because, firstly, the speed is slower than calculations without ADMM and, secondly, the accuracy is slightly decreased. Could someone please provide some suggestions on solving this? Thanks!

Regards,

Hongyang

[augusti...@chem.uzh.ch]

Dear Hongyang,

In condensed systems, there is generally less tolerance to screening. I would try lowering EPS_SCHWARZ. Note that the ADMM basis set you are currently using is meant to be used with the MOLOPT basis, which is not all-electron. For the use of ADMM in an all-electron setting, I would recommand the pcseg basis set family (and corresponding admm) by Jensen, available on Basis Set Exchange.

Best,

Augustin

----...@googlegroups.com wrote: -----

To: "cp2k" <cp...@googlegroups.com>
From: "ma455...@gmail.com"
Sent by: cp...@googlegroups.com
Date: 05/07/2021 03:00AM
Subject: [CP2K:15305] Convergence issue for GAPW without ADMM

--
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.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/c6f28c5e-6216-42ab-bab5-3f3c7b24d663n%40googlegroups.com.

[attachment "si-b-01-out.txt" removed by Augustin Bussy/at/UZH]
[attachment "si-b-01-inp.txt" removed by Augustin Bussy/at/UZH]

[ma455...@gmail.com]

Dear Augustin,

I actually have tried lowering EPS_SCHWARZ to even 1E-16, but the SCF still didn't converge. 

The calculation was run using GAPW without ADMM. I didn't use ADMM in this calculation. But it is quite weired that pob-TZVP with GAPW cannot converge. I don't know what is the reason causing instability.

Thanks,

Hongyang

[augusti...@chem.uzh.ch]

Dear Hongyang,

I tested your input file and it does actually converge, although after 322 SCF steps. The total energy oscillates along the way, which usually happens with metals and small gap semiconductors. Adding some smearing may help.

Note that you have an important HFX warning about the use of short range potentials. You should probably use a larger cell to insure your results are physically sound.

Best,

Augustin

----...@googlegroups.com wrote: -----

To: "cp2k" <cp...@googlegroups.com>
From: "ma455...@gmail.com"
Sent by: cp...@googlegroups.com

Date: 05/07/2021 11:37AM
Subject: Re: [CP2K:15307] Convergence issue for GAPW without ADMM

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/d8c52869-2a0f-432b-b87c-489844cf0979n%40googlegroups.com.

[ma455...@gmail.com]

Dear Augustin,

Sorry, I didn't run so many cycles because when I used ADMM it converged withn 20 cycles.. Yes, I noticed that WARNING too. I remembered in one tutorial lecture, it says that if HSE06 is used, then that WARNING can be ignored. 

I will try adding some smear. Thank you very much!

Regards,

Hongyang

[ma455...@gmail.com]

Dear Augustin,

I just tried adding smearing, but it still doesn't improve too much. It still takes more than 300 cycles to converge. This is really strange because the system is just pure Si and the pob-TZVP as well as HSE06 are very good basis set and functional. I don't understand why it is so hard to converge. When I use pure PBE, the convergence is really good (maybe around 20 cycles). However, HSE06 performs terrible. Do you have some suggestions in this situation?

Regards,

Hongyang

[Marcella Iannuzzi]

Dear Hongyang,

Have you tried by starting from the RESTART.wfn of the PBE calculation?

Best

marcella

[Matt W]

Yes. Runs fine for me starting from PBE guess. 

Make sure you get the electron repulsion integrals into memory - you'll get a x100+ speed up.

Matt

[Matt W]

Having said that the HSE calc is a bit on the edge. Really needs a bigger cell and EPS_SCHWARZ a bit smaller than I had above.

[ma455...@gmail.com]

Hi Marcella and Matt

Thanks! Restarting from the PBE wavefunction really helps converging a lot. The smaller EPS_SCHWARZ and EPS_SCHWARZ_FORCES decrease the calculation speed significantly. So I normally use EPS_SCHWARZ=1E-6 and EPS_SCHWARZ=1E-5. 

Regards,

Hongyang