SCAN XC density functional: Do we have to use specific PSEUDOPOTENTIALS for SCAN? Are they available?

1,166 views
Skip to first unread message

Vyacheslav Bryantsev

unread,
Apr 9, 2020, 10:50:30 AM4/9/20
to cp2k
Dear CP2K Community,

Is there a set or recommended setting for using SCAN in cp2k?

More specifically, 
1. Can we use PBE pseudopotentials for SCAN or this is a bad idea? If not, where one can find specific pseudopotentials reparametrized for SCAN?
2. Does the D3 correction work automatically now with SCAN. If not, how to specify it?
3. Recommendations for grid when using SCAN

Thank you,
Slava  


hut...@chem.uzh.ch

unread,
Apr 10, 2020, 6:31:42 AM4/10/20
to cp...@googlegroups.com
Hi

1) Many people have used PBE pp previously.
SCAN optimized pp can be found at
https://github.com/juerghutter/GTH/blob/master/SCAN/POTENTIAL
2) Newly published SCAN parameters for D3 are available in the current
version of CP2K from Github. For older version you need to
add a line in the input with the parameters.
3) Use very high cutoffs, depending on your system. If the cutoff is not
high enough SCF will not converge smoothly to a low value (10^-7 in OT).
I don't have experience if the smoothing methods work to reduce the cutoff.

regards

Juerg Hutter
--------------------------------------------------------------
Juerg Hutter Phone : ++41 44 635 4491
Institut für Chemie C FAX : ++41 44 635 6838
Universität Zürich E-mail: hut...@chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp...@googlegroups.com wrote: -----
To: "cp2k" <cp...@googlegroups.com>
From: "Vyacheslav Bryantsev"
Sent by: cp...@googlegroups.com
Date: 04/09/2020 04:50PM
Subject: [CP2K:13100] SCAN XC density functional: Do we have to use specific PSEUDOPOTENTIALS for SCAN? Are they available?
--
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/43781524-ffdc-45d0-8899-258ace32638e%40googlegroups.com.

Vyacheslav Bryantsev

unread,
Apr 13, 2020, 10:09:19 AM4/13/20
to cp2k
Dear Juerg and All,

Thank you for recommendations.
With regards to very high cutoffs, what should I try?

I startred with CUTOFF 600 and REL_CUTOFF 100, and this does not seem to work.
Which setting would you recommends for SCAN XC?

Thank you,
Slava


    &MGRID

      CUTOFF 600

      REL_CUTOFF 100

    &END MGRID

    &SCF

      MAX_SCF     150

      EPS_SCF     1.0E-6

      SCF_GUESS   RESTART

      &OUTER_SCF

        EPS_SCF 1.0E-6

        MAX_SCF 40

      &END

      &OT T

        MINIMIZER         CG

        PRECONDITIONER    FULL_ALL

      &END OT

 To unsubscribe from this group and stop receiving emails from it, send an email to cp...@googlegroups.com.

Vyacheslav Bryantsev

unread,
Apr 13, 2020, 1:55:08 PM4/13/20
to cp2k
Dear Juerg and All,

Here I provide an update for a large band gap system, consisting of Mg2+ and Cl- ions (molten salt).  
Very high CUTOFF 2300 and REL_CUTOFF 100 were necessary to get converged SCF results with SCAN-D3 XC, suitable for MD. 
Using lower cutoffs, such as  CUTOFF 2000 and REL_CUTOFF 100 were not sufficient, because it required more > 100 SCF steps for some fraction of MD steps, which is not acceptable. Because of that, the average time for AIMD was higher than for the case with CUTOFF 2300.

The portion of used input file is shown below. The average time for each SCF step is 5-6 times slower compared to PBE-D3. 
Is this the expected behavior? Are there any ways I can explore to speed up SCAN-D3 calculations, used in the context of AIMD.

Thank you,
Slava

    &MGRID

      CUTOFF 2300

      REL_CUTOFF 100

!      NGRIDS 4

    &END MGRID

    &SCF

      MAX_SCF     150

      EPS_SCF     3.5E-6

      SCF_GUESS   RESTART

      &OUTER_SCF

        EPS_SCF 3.5E-6

        MAX_SCF 40

      &END

      &OT T

        MINIMIZER         CG

        PRECONDITIONER    FULL_ALL

      &END OT

    &END SCF


     &XC

       &XC_FUNCTIONAL

         &LIBXC

          FUNCTIONAL MGGA_X_SCAN

         &END LIBXC

         &LIBXC

           FUNCTIONAL MGGA_C_SCAN

          &END LIBXC

       &END XC_FUNCTIONAL

      &vdW_POTENTIAL

        DISPERSION_FUNCTIONAL PAIR_POTENTIAL

        &PAIR_POTENTIAL

           R_CUTOFF 40.0

           TYPE DFTD3

           D3_SCALING 1.0 1.324 0.0

           PARAMETER_FILE_NAME            dftd3.dat

!           REFERENCE_FUNCTIONAL SCAN

        &END PAIR_POTENTIAL

      &END vdW_POTENTIAL

    &END XC

  &END DFT


  &SUBSYS

    &CELL

      ABC                          24.450 24.450 24.450

      PERIODIC                     XYZ

    &END CELL

    &TOPOLOGY

      COORD_FILE_FORMAT XYZ

      COORD_FILE_NAME last_frame_wrapped.xyz

    &END TOPOLOGY

    &KIND Mg

      BASIS_SET DZVP-MOLOPT-SR-GTH

      POTENTIAL GTH-SCAN-q10

    &END KIND

    &KIND Cl

      BASIS_SET DZVP-MOLOPT-SR-GTH

      POTENTIAL GTH-SCAN-q7

    &END KIND

  &END SUBSYS

Thomas Kühne

unread,
Apr 13, 2020, 2:11:55 PM4/13/20
to cp...@googlegroups.com
Dear Vyacheslav, 

such a large density cutoff strikes me as odd. SCF convergence behavior isn’t 
a good indicator to converge CUTOFF and REL_CUTOFF. Please have a look 
IMHO, Total charge density on r- and g-space grids are 
much stronger indicators …

Cheers, 
Thomas

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/6cf623c7-259a-492e-9737-a2d6105f448f%40googlegroups.com.


Vyacheslav Bryantsev

unread,
Apr 13, 2020, 3:01:31 PM4/13/20
to cp2k
Dear Thomas.

I am familiar with this tutorial. 
The purpose for having so high CUTOFF and REL_CUTOFF is not to get the converged results, but simply be able to converge SCF for consecutive MD steps.
For example,  using CUTOFF 1500 and  REL_CUTOFF  200 would give converged SCF cycles taking 1000s of steps for consecutive MD steps. Less than 10 SCF would be needed  by using CUTOFF 2300, as was empirically found. 
I will be happy to provide complete input files, if needed.

Thank you,
Slava


On Monday, April 13, 2020 at 2:11:55 PM UTC-4, tkuehne wrote:
Dear Vyacheslav, 

such a large density cutoff strikes me as odd. SCF convergence behavior isn’t 
a good indicator to converge CUTOFF and REL_CUTOFF. Please have a look 
IMHO, Total charge density on r- and g-space grids are 
much stronger indicators …

Cheers, 
Thomas

Thomas Kühne

unread,
Apr 13, 2020, 3:15:00 PM4/13/20
to cp...@googlegroups.com
Dear Slava, 

I understand. What I wanted to say that these things are only very weakly related. 
Instead of tempting around with the density cutoff values, try to improve the SCF 
section, or directly change to a diagonalization based eigenvalue solver. 
Using OT, you may try to reduce MAX_SCF to 50 for instance, reduce stepsize 
and energy_gap …

Cheers, 
Thomas

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/d6b3c9f7-bb83-40a9-88ba-77559bb2eb0f%40googlegroups.com.



==============================
Thomas D. Kühne
Dynamics of Condensed Matter
Chair of Theoretical Chemistry
University of Paderborn
Warburger Str. 100
D-33098 Paderborn
Germany

Marcella Iannuzzi

unread,
Apr 14, 2020, 3:18:46 AM4/14/20
to cp2k
Dear Slava,

Have you tried GAPW. A much lower cutoff should be sufficient and it possibly improves the convergence.
Kind regards, 
Marcella

Matt W

unread,
Apr 14, 2020, 5:41:12 PM4/14/20
to cp2k
There could be two factors to the reported huge cutoff for the MgCl2 salt:

(i) Mg is nearly pathological, you need a huge cutoff with unsmoothed GPW scheme to get the 2s semi-core states reasonalbly (so GAPW is much better)
(ii) SCAN is known to be 'difficult' in some sense, needed much tighter settings (meta GGA). I am not sure if it is compatible with GAPW method?

For other sytems the cutoff should be less extreme.

Matt

Vyacheslav Bryantsev

unread,
Apr 23, 2020, 8:36:01 AM4/23/20
to cp2k
Dear All,

Thank you for your helpful comments.
Indeed, Mg2+ is patological, because when I consider ZnCl2 melt, I have no problem with SCF convergence at much lower grid values. 
I could still get results by using a lot of CPUs. I have no experience with GAPW. Is it going to be much slower than GPW?

Thank you,
Slava

Hua Tian

unread,
Jun 4, 2020, 6:22:38 AM6/4/20
to cp...@googlegroups.com
Dear Prof. Hutter and Bryantsev,
In my FPMD simulations for aqueous solutions by using SCAN functional (with the MOLOPT-SR basis sets and the GTH  pseudopotentials for SCAN from the website in your e-mail.), I found that smoothing methods (e.g., &XC_GRID / XC_DERIV NN10_SMOOTH / XC_SMOOTH_RHO NN10 / &END XC_GRID) was probably important for the convergence of SCF. For a simulation using SCAN functional with CUTOFF 400 that SCF did not converge, adding the &XC_GRID part methoned above made the SCF converge. However, I did not know if this setting would influence the accuracy of simulations.
Sincerely, 
Tian Hua


在 2020年4月10日星期五 UTC+8下午6:31:42,jgh写道:
 To unsubscribe from this group and stop receiving emails from it, send an email to cp...@googlegroups.com.

Vyacheslav Bryantsev

unread,
Jul 26, 2020, 9:56:57 PM7/26/20
to cp2k
Dear Tian and CP2K developers,

I confirm for my system, which is a concentrated  aqueous salt, these options work with  CUTOFF=800 and it would probably work at smaller values also.

        &XC_GRID

               XC_DERIV NN10_SMOOTH

               XC_SMOOTH_RHO NN10

        &END XC_GRID
Interestingly, removing them requires CUTOFF of at least 1200.
A question to developers: What do these parameters do and what do those specific options mean?
There is very little information in the manual about these options.

Thank you
Slava
Vyacheslav Bryantsev

Hua Tian

unread,
Jul 26, 2020, 11:00:24 PM7/26/20
to cp2k
Dear Bryantsev and CP2K developers,
  In fact, I have the same question. The other day I did some tests, and found that the misconvergence of SCF could be probably attributed to the default PW method in XC_DERIV.
  According to the Reference Manual, the default parameters for the &XC_GRID section are:
&XC_GRID
  XC_DERIV PW
  XC_SMOOTH_RHO NONE
&END XC_GRID
In my FPMD simulations for aqueous solutions by using SCAN functional, with these settings, SCF did not converge even with CUTOFF 600.
  However, when I followed the guidance in www.cp2k.org/gpw, and set XC_DERIV SPLINE3, i.e., 
&XC_GRID
  XC_DERIV SPLINE3
  XC_SMOOTH_RHO NONE
&END XC_GRID
the SCF converged.
  Besides, the NN10_SMOOTH / NN10 settings I mentioned before were also applicable for CUTOFF 600.
  I am also hoping for the explanation from CP2K developers.
Sincerely,
Tian Hua


在 2020年7月27日星期一 UTC+8上午9:56:57,Vyacheslav Bryantsev写道:

Thomas Kühne

unread,
Jul 27, 2020, 1:54:21 AM7/27/20
to 'Dorothea Golze' via cp2k
Dear Vyacheslav and Tian, 

the usage of post-GGA functionals involves the calculation of \nabla n 
and nasty derivatives of the form \partial e^{XC} / \partial \nabla n, 
which gives rise to strong so-called „ringing effects“. 
To avoid the latter a number of smoothing operators such as NN10, 
SPLINE3, etc. are provided, whose numerical derivatives assume less 
continuity. Many of these smoothing operators also have a favorable 
convergence behavior wrt to the density cutoff, which otherwise would 
require rather high values to filter out the highest frequencies. 

Greetings, 
Thomas

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/044e8b24-37d9-4ec3-89bf-bc3a11185359o%40googlegroups.com.

Vyacheslav Bryantsev

unread,
Jul 29, 2020, 6:40:51 PM7/29/20
to cp2k
Thank you, Thomas,
I wanted to check that it is safe to use these non-default options for production calculations, as long as the density cutoffs are reasonable.

Slava


On Monday, July 27, 2020 at 1:54:21 AM UTC-4, tkuehne wrote:
Dear Vyacheslav and Tian, 

the usage of post-GGA functionals involves the calculation of \nabla n 
and nasty derivatives of the form \partial e^{XC} / \partial \nabla n, 
which gives rise to strong so-called „ringing effects“. 
To avoid the latter a number of smoothing operators such as NN10, 
SPLINE3, etc. are provided, whose numerical derivatives assume less 
continuity. Many of these smoothing operators also have a favorable 
convergence behavior wrt to the density cutoff, which otherwise would 
require rather high values to filter out the highest frequencies. 

Greetings, 
Thomas

Reply all
Reply to author
Forward
0 new messages