Switching from GPW to GAPW

[Léon Luntadila Lufungula]

Dear all,

I want to perform an AIM analysis with Critic2 which (for as far as I know) requires an all-electron density (i.e. not the density obtained from a GPW calculation) and also plan to calculate NMR chemical shifts in the future which will also require the all-electron GAPW method. Currently, I have figured out how to do calculations with the GPW method but I'm still a bit unsure about how to go from a GPW to a GAPW calculation. Is there someone who could elaborate on what to look at when modifying my input file? I've looked into it a couple of times, but it's a lot of different parameters and I can't seem to figure out how it works...

All the best,

Léon

[Jürg Hutter]

Hi

For a pseudopotential calculation you can just change "METHOD GPW" to
"METHOD GAPW".
If you want to do all electron calculations, you also need to use corresponding
basis sets and potentials, e.g.
BASIS_SET SVP-MOLOPT-GGA-ae
POTENTIAL ALL
Use the default settings first, check the lines
Total charge density (r-space): -0.0000000683
Total Rho_soft + Rho0_soft (g-space): -0.0000000043
in the output. If those values are larger than 0.00001 you need probably to
change defaults.
See the regtests and manual for more information.
Come back here for if this basic setup fails.

regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of Léon Luntadila Lufungula <Leon.luntad...@uantwerpen.be>
Sent: Wednesday, November 8, 2023 5:21 PM
To: cp2k
Subject: [CP2K:19481] Switching from GPW to GAPW

--
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/670d2377-8c38-4444-9130-14ebd3d32d09n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/670d2377-8c38-4444-9130-14ebd3d32d09n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Léon Luntadila Lufungula]

Dear Prof. Hutter,

Thanks for pointing me into the right direction!

Unfortunately, the defaults are not sufficient for my system as I am getting rather large values for the lines you suggested to check:

(Inner SCF loop 1)
  Total charge density (r-space):                      0.0000427891
  Total Rho_soft + Rho0_soft (g-space):         0.0001310539
(Inner SCF loop 2)
  Total charge density (r-space):                      0.0000494000
  Total Rho_soft + Rho0_soft (g-space):         0.0001376711

I have attached my input file and the short part that was written to the output if you would like to see it in more detail. My system consists of a TiO2 [anatase (101)] slab with organic molecules adsorbed onto the surface, so I suspect that it might be the Ti atoms for which the defaults are insufficient.

I have been looking into GAPW in the regtests, documentation and forum over the last couple of days and from what I understand, the most important parameters to change should be the following:
(FORCE_EVAL / SUBSYS / KIND)

• LEBEDEV_GRID: I'm a bit unsure how to get a correct value for this parameter for each atom kind. Does this involve a convergence test for each atom type individually or is there some rule of thumb?
• RADIAL_GRID: same as for LEBEDEV_GRID.

(FORCE_EVAL /DFT / QS)

• EPS_DEFAULT: This one is already at 1.0E-12, should I reduce it further?
• EPS_FIT: The default is 1.0E-4, but I've read in another post in a reply from M. Krack that it is too large for most systems and should be reduced to 1.0E-6. Should I reduce it in my case?
• EPS_ISO: I've seen it mentioned a couple of times, but never if the value should be reduced or what to do with it exactly.
• EPS_RHO0: According to the reply of M. Krack smaller values than the default give better accuracy.

(FORCE_EVAL / DFT / MGRID)

• CUTOFF: This is generally lower than for a GPW calculation as I understand it but I guess I should perform a convergence test to choose an appropriate value. Alternatively, I could also multiply the relative cutoff by the maximum exponent in the soft basis set (PRINT_LEVEL MEDIUM to get the values right?) as you and Marcella suggested in another post.
• REL_CUTOFF: I'm a bit unsure about how to choose this as all the convergence schemes I've read about first converge the CUTOFF with a large REL_CUTOFF and then reduce the value of REL_CUTOFF to get the best balance between accuracy and efficiency, but as explained above, the rule of thumb for the cutoff uses the value of REL_CUTOFF to determine the CUTOFF...

Any suggestions on how to optimize these parameters would be greatly appreciated!

I also have two other quick questions about your reply:

1. I thought that GPW was used for pseudopotential calculations and GAPW was used for all electron calculations, but now you say that you can also do pseudopotential calculations with GAPW. Is there an advantage to using GAPW over GPW? I know that the disadvantage is that some features are not present for GAPW calculations, but is there a reason to use GPW if all your intended features are implemented for GAPW?
2. The BASIS_MOLOPT_UZH file contains SVP,- TZVPP- and QZVPP-MOLOPT-GGA-ae basis sets. Is there also a double zeta all electron (MOLOPT) basis set somewhere I can use? I have always been told that single zetas are not intended for production runs, but I suspect that TZVPP is already quite computationally expensive for a pre-optimization.

My apologies for the barrage of questions, but I don't have anyone in my group with experience in CP2K so it's hard sometimes to figure these things out on my own...

Thanks in advance and for your previous reply!

Kind regards,

Léon

[Marcella Iannuzzi]

Dear Leon,

Yes you should probably use better basis sets. 

The Basis Set Exchange at https://www.basissetexchange.org offers a range of suitable options, some of which are already available in the EMSL_BASIS_SETS file.

One advantage of GAPW is the reduced PW cutoff necessary 

Regards

Marcella

[Jürg Hutter]

Hi

you can add
RADIAL_GRID 100
LEBEDEV_GRID 100
HARD_EXP_RADIUS 2.10
to Ti KIND section
This is my QS section for accurate calculations
&QS
METHOD GAPW
EPS_DEFAULT 1.E-14
EPSFIT 1.E-4
EPSISO 1.E-12
EPSRHO0 1.E-6
LMAXN0 4
GAPW_1C_BASIS EXT_SMALL
&END QS
The save side for the cutoff would be
&MGRID
CUTOFF 400
REL_CUTOFF 60
&END MGRID

Yes, you can perform pseudopotential calculations with GAPW. You can also
mix all electron and pseudopotentials.

Please note that the all electron basis sets in BASIS_MOLOPT_UZH are modeled
after the def2 Ahlrichs basis sets. In their nomenclature SVP stands for
split-valence-polarisation. So this is equivalent to a DZVP basis from other schools (double zeta valence + polarisation) and not a single zeta basis.
You can easily verify this by checking the number of functions.

regards
JH

________________________________________
From: cp...@googlegroups.com <cp...@googlegroups.com> on behalf of Léon Luntadila Lufungula <Leon.luntad...@uantwerpen.be>

Sent: Tuesday, November 14, 2023 12:04 PM
To: cp2k
Subject: Re: [CP2K:19498] Switching from GPW to GAPW

Dear Prof. Hutter,

Thanks for pointing me into the right direction!

Unfortunately, the defaults are not sufficient for my system as I am getting rather large values for the lines you suggested to check:

(Inner SCF loop 1)
Total charge density (r-space): 0.0000427891
Total Rho_soft + Rho0_soft (g-space): 0.0001310539
(Inner SCF loop 2)
Total charge density (r-space): 0.0000494000
Total Rho_soft + Rho0_soft (g-space): 0.0001376711

I have attached my input file and the short part that was written to the output if you would like to see it in more detail. My system consists of a TiO2 [anatase (101)] slab with organic molecules adsorbed onto the surface, so I suspect that it might be the Ti atoms for which the defaults are insufficient.

I have been looking into GAPW in the regtests, documentation and forum over the last couple of days and from what I understand, the most important parameters to change should be the following:
(FORCE_EVAL / SUBSYS / KIND)

* LEBEDEV_GRID: I'm a bit unsure how to get a correct value for this parameter for each atom kind. Does this involve a convergence test for each atom type individually or is there some rule of thumb?
* RADIAL_GRID: same as for LEBEDEV_GRID.

(FORCE_EVAL /DFT / QS)

* EPS_DEFAULT: This one is already at 1.0E-12, should I reduce it further?
* EPS_FIT: The default is 1.0E-4, but I've read in another post in a reply from M. Krack<https://groups.google.com/g/cp2k/c/FUvjXq9RvWY/m/8VwHvHGUAgAJ> that it is too large for most systems and should be reduced to 1.0E-6. Should I reduce it in my case?
* EPS_ISO: I've seen it mentioned a couple of times, but never if the value should be reduced or what to do with it exactly.
* EPS_RHO0: According to the reply of M. Krack<https://groups.google.com/g/cp2k/c/FUvjXq9RvWY/m/8VwHvHGUAgAJ> smaller values than the default give better accuracy.

(FORCE_EVAL / DFT / MGRID)

* CUTOFF: This is generally lower than for a GPW calculation as I understand it but I guess I should perform a convergence test to choose an appropriate value. Alternatively, I could also multiply the relative cutoff by the maximum exponent in the soft basis set (PRINT_LEVEL MEDIUM to get the values right?) as you and Marcella suggested in another post<https://groups.google.com/g/cp2k/c/JX8bDi_w478/m/okVfMy6EAQAJ>.
* REL_CUTOFF: I'm a bit unsure about how to choose this as all the convergence schemes I've read about first converge the CUTOFF with a large REL_CUTOFF and then reduce the value of REL_CUTOFF to get the best balance between accuracy and efficiency, but as explained above, the rule of thumb for the cutoff uses the value of REL_CUTOFF to determine the CUTOFF...

Any suggestions on how to optimize these parameters would be greatly appreciated!

I also have two other quick questions about your reply:

1. I thought that GPW was used for pseudopotential calculations and GAPW was used for all electron calculations, but now you say that you can also do pseudopotential calculations with GAPW. Is there an advantage to using GAPW over GPW? I know that the disadvantage is that some features are not present for GAPW calculations, but is there a reason to use GPW if all your intended features are implemented for GAPW?
2. The BASIS_MOLOPT_UZH file contains SVP,- TZVPP- and QZVPP-MOLOPT-GGA-ae basis sets. Is there also a double zeta all electron (MOLOPT) basis set somewhere I can use? I have always been told that single zetas are not intended for production runs, but I suspect that TZVPP is already quite computationally expensive for a pre-optimization.

To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/36846b0d-2a8d-4ce1-a35e-991b76854170n%40googlegroups.com<https://groups.google.com/d/msgid/cp2k/36846b0d-2a8d-4ce1-a35e-991b76854170n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Léon Luntadila Lufungula]

Hi Marcella and Prof. Hutter,

Thanks a lot for your valuable insights and suggestions!

Marcella, I'll definitely explore the Basis Set Exchange and check out the EMSL_BASIS_SETS file to find a more suitable basis set for my calculations.

Prof. Hutter, I gave your input a shot with the TZVPP-MOLOPT-GGA-ae basis set on an alternate HPC cluster since our usual setup is under maintenance until after the weekend. The CP2K version here is 2023.1, so I couldn't try the "GAPW_1C_BASIS EXT_SMALL" keyword as this was introduced in version 2023.2. Unfortunately, the calculation failed with a "Cholesky decomposition failed. Matrix ill-conditioned" error. I attached the input and output files for your reference.

Any guidance on troubleshooting this problem would be much appreciated.

Also, any advice on optimizing the LEBEDEV_GRID and RADIAL_GRID parameters for different atom kinds would be fantastic.

Thanks again for all the help!

Best regards,
Léon