[CP2K-user] [CP2K:15205] DFT+U unphysical warning

Nicholas Winner nwi... at berkeley.edu
Mon Apr 26 01:12:22 UTC 2021


Matthias, that all makes a lot of sense. I think I will do a few formation 
energy comparisons at the DZV / TZV levels for my own sanity, but that's 
for me to worry about.

The one thing that remains is the issue of Lowdin vs Mulliken. I still have 
some TMOs that struggle to converge even once I switch to the SR basis 
sets, but work when using Lowdin+SR basis sets.  Of course, Lowdin does not 
have forces implemented. Do you happen to know why they have yet to be 
implemented? Based on my knowledge, Lowdin and Mulliken are just two 
different ways to get q_{mu, nu}, and so I don't see why Lowdin would not 
be implemented and Mulliken is. If it is just a matter of effort, I might 
fork Cp2k and take a look at the source, but if there is a theoretical 
reason for it, then I won't bother.

Thanks for your help,
Nick
On Thursday, April 22, 2021 at 1:28:41 PM UTC-7 Matthias Krack wrote:

> Nick, we are talking about dense solid state systems. For such system, 
> space is already well covered by a DZV basis set at each atom. The 
> situation is, however, different already for molecular liquids like water 
> or for systems with larger voids (MOFs, surfaces) or even small isolated 
> molecular systems. Soft functions are required to describe the electron 
> density decaying into the void regions. More polarisation function (second 
> set of d and a set of functions) certainly improve the description, but 
>  these function sets increase the computational costs significantly while 
> their impact is rather moderate for DFT applied for O in dense solid state 
> systems. I tried TZV/QZV and/or more polarisation functions for O in such 
> systems, but the gain was rather small compared to the additional 
> computational costs. There are tutorials 
> <https://www.cp2k.org/exercises:2015_cecam_tutorial:basis_set_optimisation_using_optimize_basis?s%5b%5d=molopt&s%5b%5d=na> 
> showing how you may generate new or augment existing MOLOPT basis sets. It 
> is also important to employ a balanced set of basis sets for a system which 
> is, of course, also true for other system types. A mixing of SR and non-SR 
> basis sets should always be done with care. The optimisation of the SR 
> basis sets includes also the condition number of the overlap matrix as a 
> parameter during the basis set generation procedure which allows for a 
> control of its numerical stability.
>
>  
>
> Matthias
>
>  
>
> *Von:* c... at googlegroups.com <c... at googlegroups.com> *Im Auftrag von *Nicholas 
> Winner
> *Gesendet:* Donnerstag, 22. April 2021 20:34
> *An:* cp2k <c... at googlegroups.com>
> *Betreff:* Re: [CP2K:15205] DFT+U unphysical warning
>
>  
>
>
> Matthias, after investigating further, it does seem like using the SR 
> potential is fairly consistent in giving good results for the DFT+U 
> calculations, even when Mulliken leads to a negative +U energy. I have 
> previously not used DFT+U, and always used TZV2PX for O finding no 
> problems, but the population analysis does seem to require a better 
> conditioned overlap matrix. 
>
>  
>
> Matt mentioned TZV-SR basis sets in the UCL file, which is true, but not 
> for O. In fact there are a few elements which only go up to DZVP-SR across 
> all the BASIIS_MOLOPT_{, UCL, LnPP1} files. Some of these are O, C, Cl, F, 
> H, etc. which are very common elements. H for instance only has SZV-SR and 
> the rest are non-SR. If I believe that SR should always be used for 
> stability in condensed phase, then it seems like these need some new basis 
> sets, at least up to TZV quality. Maybe you disagree and DZVP is sufficient 
> for these elements for some reason, while others really need larger bases?
>
>  
>
> Is there a reference for the procedure for fitting the SR basis sets? The 
> molopt sets are described in "Joost VandeVondele and Juerg Hutter, J. Chem. 
> Phys. 127, 114105 (2007)", but the BASIS_MOLOPT file simply says "variants 
> of these basis sets using less and thus less diffuse primitives " were used 
> for the SR versions, which is a little bit vague to me.
>
> On Thursday, April 22, 2021 at 2:18:31 AM UTC-7 Matt W wrote:
>
> Just to note there are MOLOPT-TZVP-SR in the BASIS_MOLOPT_UCL files 
> distributed with CP2K.
>
> Matt
>
> On Wednesday, April 21, 2021 at 11:58:19 PM UTC+1 n... at berkeley.edu 
> wrote:
>
> So something very strange happened.
>
>  
>
> I'm hesitant to switch O to a SR basis, because only SZV/DZVP are 
> available. SZV isn't production quality and DZVP really isn't near the 
> basis set limit, but I gave it a try to see what would happen. The warning 
> about negative DFT+U persists, but the energy changed to 
>  -15383.653493175701442. I then re-initialized with TZV2PX and converged to 
>  -15385.210534637011733. 
>
>  
>
> So to summarize I have three values for U=1.75
>
> -15378.643602015281431 with TZV2PX
>
> -15383.653493175701442 with DZVP
>
> -15385.210534637011733 after restarting TZV2PX from the DZVP wfn.
>
>  
>
> These are enormous differences in energy and I don't know what to make of 
> it. Any more ideas about what is going wrong based on my inputs?
>
>  
>
> On Wednesday, April 21, 2021 at 11:12:25 AM UTC-7 Matthias Krack wrote:
>
> At a first glance, the basis set TZV2PX-MOLOPT-GTH-q6 for O is not well 
> suited for condensed phase systems. You should rather use MOLOPT basis sets 
> with “SR” in the name.
>
>  
>
> HTH
>
>  
>
> Matthias
>
>  
>
> *Von:* c... at googlegroups.com <c... at googlegroups.com> *Im Auftrag von *Nicholas 
> Winner
> *Gesendet:* Mittwoch, 21. April 2021 19:36
> *An:* cp2k <c... at googlegroups.com>
> *Betreff:* Re: [CP2K:15178] DFT+U unphysical warning
>
>  
>
> Here is my input file for Mn3O4. 
>
> On Wednesday, April 21, 2021 at 10:30:13 AM UTC-7 Matthias Krack wrote:
>
> Hello Nick
>
>  
>
> The selection of the initial guess can help to achieve convergence. Could 
> you provide a case which fails to converge with MULLIKEN? Otherwise, it is 
> difficult to give further hints.
>
>  
>
> Matthias
>
>  
>
> *Von:* c... at googlegroups.com <c... at googlegroups.com> *Im Auftrag von *Nicholas 
> Winner
> *Gesendet:* Mittwoch, 21. April 2021 19:15
> *An:* cp2k <c... at googlegroups.com>
> *Betreff:* Re: [CP2K:15176] DFT+U unphysical warning
>
>  
>
> Thanks for the reply Matthias. I've had moderate success following what 
> you suggested.
>
>  
>
> One issue that still persists, however, is that for some calculations I 
> can only get convergence by using PLUS_U_METHOD LOWDIN. For example, MnO 
> with Ueff=1.75eV, diagonalization and small smearing (T=100). I've tried 
> using diagonalization and OT with MULLIKEN to get convergence, but 
> sometimes LOWDIN seems to be the only way. Unfortunately, Lowdin does not 
> support forces, so relaxations are out of the question. 
>
>  
>
> Do you have a suggestion for how to work around this?
>
>  
>
> -Nick
>
> On Friday, April 16, 2021 at 11:00:09 PM UTC-7 Matthias Krack wrote:
>
> Hello Nick
>
>  
>
> I do not recommend the use of the PLUS_U_METHOD MULLIKEN_CHARGES. Just use 
> the PLUS_U_METHOD MULLIKEN. The Mulliken population analysis can give 
> unphysical orbital occupations, i.e. values greater than one (UKS case) or 
> two (RKS case). Often the maximum occupation is only slightly exceeded and 
> the warning can be ignored safely. You can print the orbital occupations 
> for the orbitals affected by the +U correction with this print key 
> <https://manual.cp2k.org/cp2k-8_1-branch/CP2K_INPUT/FORCE_EVAL/DFT/PRINT/PLUS_U.html> 
> at the PRINT_LEVEL medium to check the actual occupation. Note, that the U 
> values found appropriate with PW codes in the literature are not 
> necessarily optimal for CP2K, too. CP2K often gives a similar effect, e.g. 
> impact on the band gap, for smaller U values.
>
>  
>
> HTH
>
>  
>
> Matthias
>
>  
>
> *Von:* c... at googlegroups.com <c... at googlegroups.com> *Im Auftrag von *Nicholas 
> Winner
> *Gesendet:* Freitag, 16. April 2021 22:57
> *An:* cp2k <c... at googlegroups.com>
> *Betreff:* [CP2K:15147] DFT+U unphysical warning
>
>  
>
> Hello all, I am running some DFT+U calculations on Mn-O systems. While I 
> have found a cp2k effective U value in the literature of ~1.3 eV for Mn. I 
> notice that when tuning the value myself, I begin seeing the following 
> warning once the U value reaches 0.5eV.
>
> *** WARNING in dft_plus_u.F:2006 :: DFT+U energy contibution is negative 
> ***
>
>  *** possibly due to unphysical Mulliken charges!    
>
> Now this is only a warning, not an indication that the calculation is 
> *necessarily* wrong, but it is troubling at least. Especially when my U 
> value is nowhere near the size of lit value. I am using PLUS_U_METHOD 
> MULLIKEN_CHARGES in order to have a marginally more robust solution. Does 
> anyone have experience with how seriously to take this warning? I don't 
> have a frame of reference to know if I should ignore it.
>
>  
>
> -Nick
>
> -- 
> 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 cp... at googlegroups.com.
> To view this discussion on the web visit 
> https://groups.google.com/d/msgid/cp2k/f41c9258-21be-4389-a69f-55aab2a808d4n%40googlegroups.com 
> <https://groups.google.com/d/msgid/cp2k/f41c9258-21be-4389-a69f-55aab2a808d4n%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 cp... at googlegroups.com.
>
> To view this discussion on the web visit 
> https://groups.google.com/d/msgid/cp2k/66f4854a-2810-4a94-a5d7-6372c2103c4cn%40googlegroups.com 
> <https://groups.google.com/d/msgid/cp2k/66f4854a-2810-4a94-a5d7-6372c2103c4cn%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 cp... at googlegroups.com.
>
> To view this discussion on the web visit 
> https://groups.google.com/d/msgid/cp2k/a7c40bdf-fdfb-4448-a20e-bb24a88a7ee9n%40googlegroups.com 
> <https://groups.google.com/d/msgid/cp2k/a7c40bdf-fdfb-4448-a20e-bb24a88a7ee9n%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 cp... at googlegroups.com.
>
> To view this discussion on the web visit 
> https://groups.google.com/d/msgid/cp2k/1f67b87b-7768-493b-b4e3-e5d46d426847n%40googlegroups.com 
> <https://groups.google.com/d/msgid/cp2k/1f67b87b-7768-493b-b4e3-e5d46d426847n%40googlegroups.com?utm_medium=email&utm_source=footer>
> .
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20210425/2a356a78/attachment.htm>


More information about the CP2K-user mailing list