[CP2K-user] [CP2K:21199] Re: Counterpoise correction with Grimme D4

[Holger Sassnick]

Hi Prof Hutter,

in principle metalorganic frameworks with water adsorbed in the pores (like 
the attached input file of the initial post). The BSSE here is around 0.05 
eV for one water molecule in the MOF. I am planning to train MLPs on the 
total energies that are in turn used in Monte Carlo simulations and the 
binding energy (MOF/H2O - MOF - H2O) is plugged into the criteria to accept 
or reject a deletion/insertion move. So I think I really need to ensure 
that the BSSE is very low in order to not bias the MC simulations.

Taking that into account, I see two options:
* Remove the BSSE via the Counterpoise correction on the whole training 
set. This would mean a multitude of calculations for each training point.
* Move to a plane-wave code.

I very much like CP2K since it is a very efficient code with extensive 
functionality. Thus, I was wondering whether it would be possible to have a 
basis set with negligible BSSE in CP2K.

All the best,
Holger

Jürg Hutter schrieb am Donnerstag, 20. Februar 2025 um 10:30:23 UTC+1:

> Hi
> What type of system and BSSE values are you interested in? There is
> a discussion on BSSE for Molopt basis sets in the original paper.
> Results are for small molecule dimers.
> See
> J. Chem. Phys. 127, 114105 (2007) https://doi.org/10.1063/1.2770708
> Section III D, Table VI
> Molopt basis sets outperform normal Gaussian basis sets. However,
> I don't know how the numerical basis sets look for these systems.
>
> regards
> JH
>
> ________________________________________
> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of Holger 
> Sassnick <holger.... at gmail.com>
> Sent: Thursday, February 20, 2025 7:54 AM
> To: cp2k
> Subject: Re: [CP2K:21177] Re: Counterpoise correction with Grimme D4
>
> Hi Prof Hutter,
>
> Thank you for the reply and insights. Unfortunately, I fear that the 
> constant BSSE would still bias the sampling if the "binding energy" is part 
> of the acceptance criteria to insert/delete a molecule in a MC scheme.
> In FHI-Aims, (from my experience) even the larger basis sets seem to 
> perform satisfactorily on molecular as well as dense systems. Do you think 
> it would be worth a try to design, e.g., a QZVPP basis set to push the 
> boundaries a bit?
> I am not sure if that would be large enough to keep the BSSE on an 
> acceptable level.
>
> All the best,
> Holger
> Jürg Hutter schrieb am Mittwoch, 19. Februar 2025 um 11:58:30 UTC+1:
> Hi
>
> the MOLOPT basis sets are optimized to perform for a wide range of
> systems (molecules, solutes, solids, open structures), while keeping a
> small number of functions for performance reasons.
> BSSE is of similar size as for comparable size Gaussian basis sets.
> For MD calclulation we tested and confirm that BSSE is typically
> a constant shift on energies and does not affect dynamics.
> If you want to have smaller/no BSSE you have to increase the
> number of functions and e.g. include the atomic ground
> state functions in the Basis (as in FHI-AIMS). There is the danger
> that these near complete basis sets will perform poorly in dense
> systems due to bad condition numbers.
>
> regards
> JH
>
> ________________________________________
> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of Holger 
> Sassnick <holger.... at gmail.com>
> Sent: Thursday, February 13, 2025 11:14 AM
> To: cp2k
> Subject: Re: [CP2K:21142] Re: Counterpoise correction with Grimme D4
>
> Hi Prof Hutter,
>
> thank you for transferring the issue to the github repository. I was also 
> wondering whether it would be reasonable and feasible to design a larger 
> MOLOPT basis set with negligible BSSE. In this context I am also thinking 
> about the training of MLPs, e.g. in this publication (
> https://pubs.acs.org/doi/10.1021/jacs.4c15287) trajectories were 
> re-calculated with VASP for the training to ensure accurate results.
>
> Fox example, in FHI-Aims the "tight" basis set parameters have hardly any 
> BSSE. I know that the code uses a different localized basis set, so maybe 
> this kind of accuracy cannot be reached with GTOs? I would be happy to hear 
> your opinion on this matter.
>
> All the best,
> Holger
>
>
> Jürg Hutter schrieb am Montag, 10. Februar 2025 um 09:39:16 UTC+1:
> Hi
>
> thank you for bringing this to our attention. I have opened a bug report
> on the CP2K github page.
> https://github.com/cp2k/cp2k/issues
> [CP2K:21126] Counterpoise correction with Grimme D4
>
> regards
> JH
>
> ________________________________________
> From: cp... at googlegroups.com <cp... at googlegroups.com> on behalf of Holger 
> Sassnick <holger.... at gmail.com>
> Sent: Monday, February 10, 2025 7:19 AM
> To: cp2k
> Subject: [CP2K:21130] Re: Counterpoise correction with Grimme D4
>
> Hi Quentin,
>
> Thank you for the help and the swift response. You are right, the 
> dispersion interaction of "ghost" atoms is (wrongly) added which leads to 
> these unreasonable values. As far as I understand, I can just manually sum 
> up the energy contributions of the fragments, replacing the dispersion 
> interaction energy with the one from the fragment without ghost atoms. 
> Taking those total energy and calculating the interaction energy and 
> BSSE-corrected total energy leads then to values similar to PBE or PBE-D3.
>
> I think it would be also great if the implementation in CP2K could take 
> care of ghost atoms and calculate all the fragment energies correctly as is 
> done for the Grimme-D3 and other van der Waals corrections.
>
> Have a great start to the week.
> All the best,
> Holger
>
> Quentin Pessemesse schrieb am Sonntag, 9. Februar 2025 um 01:54:45 UTC+1:
> Also, the order of magnitude of your "BSSE-corrected" interaction energy 
> looks like the order of magnitude of E(disp)AB, which would make sense if 
> the ghost atoms are not treated as ghost by the empirical dispersion 
> correction
>
> Le dimanche 9 février 2025 à 01:39:46 UTC+1, Quentin Pessemesse a écrit :
> Hi Holger,
> Counterpoise correction is a correction to the electronic energy, you 
> should reason only on the electronic energies. The dispersion energy of the 
> isolated fragments has no physical meaning, and the D3/D4 empirical 
> correctoin plays no part in the BSSE as it is added after the SCF, and only 
> depends on the geometry.
> Here, I think CP2k is giving dispersion correction to the ghost atoms, 
> calculating both the error that comes from basis functions of fragment 
> being used to build the density of fragment B (BSSE), and adding some 
> dispersion energy of the fragments and total system as well.
> Instead of :
> BSSE = (EA − EA(B)) + (EB − E (A)B)
> You get:
> ([E(elec)A + E(disp)A] - [E(elec)A(B) - E(disp)AB]) + ([E(elec)B + 
> E(disp)B] - [E(elec)(A)B - E(disp)AB]) = BSSE + (E(disp)A + E(disp)B - 
> 2*E(disp)AB)
> Try to run the BSSE single point calculation without the empirical 
> dispersion and substract it from the interaction energy you get with D3/D3 
> correction. If it does not fix the issue, maybe I'm mistaken and something 
> else is wrong.
> Hope this helped :)
> Q.
> Le samedi 8 février 2025 à 13:36:41 UTC+1, Holger Sassnick a écrit :
> Hello,
>
> lately I have been trying to calculate the adsorption energy of water in 
> the CAU-23 MOF with different XC functionals. To avoid the quite 
> significant BSSE of CP2K's MOLOPT basis sets, I had to apply a counterpoise 
> correction.
>
> However, when using the Grimme D4 method the obtained values didn't really 
> make sense as they resulted in a positive interaction energy (I have 
> attached the corresponding input and output files). The same calculation 
> with the PBE functional or PBE + Grimme D3 gives a negative interaction 
> energy (the absolute value is also significantly smaller).
>
> I was wondering whether this is a bug in the code? Would it be possible 
> that the ghost atoms are not properly treated by the interface to the DFTD4 
> library?
>
> I would be very grateful for some feedback.
>
> Thanks in advance and all the best,
> Holger
>
> --
> 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... at googlegroups.com<mailto:cp2k+uns... at googlegroups.com
> >.
> To view this discussion visit 
> https://groups.google.com/d/msgid/cp2k/f83527a5-d511-46cf-9ce3-19cdcb4ffbc5n%40googlegroups.com
> <
> https://groups.google.com/d/msgid/cp2k/f83527a5-d511-46cf-9ce3-19cdcb4ffbc5n%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+uns... at googlegroups.com<mailto:cp2k+uns... at googlegroups.com
> >.
> To view this discussion visit 
> https://groups.google.com/d/msgid/cp2k/064e690d-8716-4fe4-881e-9e02ff958213n%40googlegroups.com
> <
> https://groups.google.com/d/msgid/cp2k/064e690d-8716-4fe4-881e-9e02ff958213n%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+uns... at googlegroups.com<mailto:cp2k+uns... at googlegroups.com
> >.
> To view this discussion visit 
> https://groups.google.com/d/msgid/cp2k/0006b840-eaeb-48a9-8eab-0c0f829fede3n%40googlegroups.com
> <
> https://groups.google.com/d/msgid/cp2k/0006b840-eaeb-48a9-8eab-0c0f829fede3n%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 visit https://groups.google.com/d/msgid/cp2k/440055ec-b7ab-4524-8d50-7f332a168f84n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20250223/8979e496/attachment-0001.htm>