[CP2K-user] [CP2K:10804] Re: Problem in the preconditioner when using UKS

[Daniele Ongari]

Dear Dr. Krack,
for the Fe-MOF-74 case I tried already FULL_KINETIC but it didn't work, so 
I neglected that any preconditioner which is assumed to be less robust than 
FULL_ALL would work. I will test the different ones. 
I agree that a black-box approach is not generally a good idea for these 
cases but practically speaking this 2 step protocol that we are using is 
working well, and to give an idea, out of ca. 300 frameworks (MOFs, COFs 
and zeolites), less then 10% passed to to the second step (i.e., CG + no 
outer SCF). So, I'm thinking if I can improve the efficiency even more but 
I'm already happy with these results!

Daniele

Il giorno giovedì 4 ottobre 2018 15:10:30 UTC+2, Matthias Krack ha scritto:
>
> Dear Daniele
>
>  
>
> I am not aware of any control key to stop an SCF run as soon as the 
> electron count becomes unreasonable. It shouldn’t be too difficult, 
> however, to implement such a control. The use of an outer SCF should 
> converge faster in most cases. Did you try already a different 
> preconditioner like FULL_SINGLE_INVERSE with OT CG and an outer SCF? If 
> such choice does not result in a robust convergence behaviour then you are 
> dealing with problematic systems which you can hardly run in a black-box 
> manner.
>
>  
>
> Best,
>
>  
>
> Matthias
>
>  
>
> *From:* cp... at googlegroups.com <javascript:> <cp... at googlegroups.com 
> <javascript:>> *On Behalf Of *Daniele Ongari
> *Sent:* Donnerstag, 4. Oktober 2018 14:47
> *To:* cp2k <cp... at googlegroups.com <javascript:>>
> *Subject:* Re: [CP2K:10804] Re: Problem in the preconditioner when using 
> UKS
>
>  
>
> Thank you a lot!
> This was a very good hint. Indeed, it is clear from the  "Total charge 
> density on r-space grids" that something is going weird with the count of 
> electron as soon as the energy starts to diverge!
>
> *$egrep 'outer SCF|Total charge density on r-space grids' diis.out*
> diis.out:  Total charge density on r-space grids:        0.0000001146
> diis.out:  outer SCF iter =    1 RMS gradient =   0.47E-04 energy =      
> -1301.1851926636
> diis.out:  Total charge density on r-space grids:        0.0000001146
> diis.out:  outer SCF iter =    2 RMS gradient =   0.42E-04 energy =      
> -1301.2195780393
> diis.out:  Total charge density on r-space grids:        0.0000001047
> diis.out:  outer SCF iter =    3 RMS gradient =   0.14E-05 energy =      
> -1301.2275937198
> diis.out:  Total charge density on r-space grids:        0.0003332641
> diis.out:  outer SCF iter =    4 RMS gradient =   0.62E-05 energy =      
> -1301.2754195817
> diis.out:  Total charge density on r-space grids:       -0.0424028684
> diis.out:  outer SCF iter =    5 RMS gradient =   0.96E-04 energy =      
> -1318.9753098223
> diis.out:  Total charge density on r-space grids:        0.5522536311
> diis.out:  outer SCF iter =    6 RMS gradient =   0.13E-03 energy =      
> -1317.2906799313
> diis.out:  Total charge density on r-space grids:        3.4918015156
> diis.out:  outer SCF iter =    7 RMS gradient =   0.52E-03 energy =      
> -1315.3866686940
> diis.out:  Total charge density on r-space grids:        4.3723629910
> diis.out:  outer SCF iter =    8 RMS gradient =   0.97E-03 energy =      
> -1298.6196710054
> diis.out:  Total charge density on r-space grids:        9.1128313578
> diis.out:  outer SCF iter =    9 RMS gradient =   0.20E-02 energy =      
> -1284.6336624946
> diis.out:  Total charge density on r-space grids:        8.4010321552
> diis.out:  outer SCF iter =   10 RMS gradient =   0.20E-02 energy =      
> -1278.9516795867
> diis.out:  Total charge density on r-space grids:        7.2155551466
> diis.out:  outer SCF iter =   11 RMS gradient =   0.20E-02 energy =      
> -1285.4339788964
> diis.out:  outer SCF loop FAILED to converge after   11 iterations or  550 
> steps
>
>  
>
> I'm now wondering: is there a way to stop CP2K as soon as the  "Total 
> charge density on r-space grids"  exceeds a threshold (let's say 1.0) to 
> avoid wasting time and switch to the CG algorithm without outer SCF?
> I'm using this 2 step protocol for high-throughput DFT calculations, 
> therefore it would be helpful not to waste time!
>
> Regards,
> Daniele
>
>  
>
>
> Il giorno mercoledì 3 ottobre 2018 21:45:20 UTC+2, Matthias Krack ha 
> scritto:
>
> Dear Daniele
>
>  
>
> You should take note of the electron count in the OT DIIS output and ask 
> yourself, if the corresponding energy has any meaning. 
>
> The OT CG run converged to some state at least.
>
>  
>
> HTH
>
>  
>
> Matthias
>
>  
>
> *Von:* cp... at googlegroups.com <cp... at googlegroups.com> *Im Auftrag von *Daniele 
> Ongari
> *Gesendet:* Mittwoch, 3. Oktober 2018 19:47
> *An:* cp2k <cp... at googlegroups.com>
> *Betreff:* [CP2K:10799] Re: Problem in the preconditioner when using UKS
>
>  
>
> [image: Screenshot from 2018-10-03 19-42-55.png]
>
> Hi,
> here I come with a similar problem for Co-MOF-74 using different settings.
> In the legend of the figure:
> "diis" is using OT DIIS with 50 inner and 10 outer SCF cycles
> "cg" is using OT CG with 2000 inner and no outer SCF cycles
>
>
> I think the problem is related to the bandgap going close to zero and 
> making the preconditioner to diverge, but this time the energy assumes even 
> lower energies than the final ones. 
> Does it makes sense to have a lower energy being DFT not strictly 
> variational, and this should be seen as a merely mathematical mess?
>
> The calculation is converging when I run the OT CG minimization without 
> outer steps: can I consider the final result of this calculation as 
> reliable?
>
> Thanks
>
> Daniele
>
>  
>
>
> Il giorno mercoledì 19 settembre 2018 12:50:34 UTC+2, Daniele Ongari ha 
> scritto:
>
> Dear CP2K developers, 
> I want to report a serious problem with the preconditioner when computing 
> the energy of Fe-MOF-74 with UKS settings.
> Long story short the energy goes down, it is close to convergence but 
> then, after recomputing the preconditioner in the outer step it starts to 
> diverge: see the image.
>
>  
>
> [image: image.png]
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
> Now, I attached the input and the output, but I already tried very 
> different settings:
> -smearing
> -diagonalization without smearing
> -different BS
> -UKS false (converged, but to a higher energy, forcing the closed shell)
> - starting magnetization on metals
> - CG minimizer
> - lower multiplicity
> - 2x and 3x replicated unit cell in the shortest dimension
> - lowering ENERGY_GAP to 0.001
> These calculation ALL failed (I made 16 attempts in total, changing 
> settings!).
>
> And at the end the only one working (the 17th attempt, ironically the 
> unlucky number for Italians!) was to use CG minimizer with *no outer SCF*, 
> hence not recomputing the preconditioner. It converged after ~600 inner 
> steps, to:
>
>  
>
>  HOMO - LUMO gap [eV] :    1.917671
>
>  HOMO - LUMO gap [eV] :    0.235602
>
>  
>
>  ENERGY| Total FORCE_EVAL ( QS ) energy (a.u.):            
> -1170.045288008604075
>
>  
>
> Please let me know what was the problem, if it is a known issue of the 
> preconditioner and how to circumvent it: for the moment I'm using no outer 
> steps to have a robust convergence, but I know that this may cause the 
> convergence to a local miniimum!
>
> Thanks a lot!
>
> Daniele 
> PhD candidate, LSMO, EPFL Sion
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
>  
>
> -- 
> 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+... at googlegroups.com.
> To post to this group, send email to cp... at googlegroups.com.
> Visit this group at https://groups.google.com/group/cp2k.
> For more options, visit https://groups.google.com/d/optout.
>
> -- 
> 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+... at googlegroups.com <javascript:>.
> To post to this group, send email to cp... at googlegroups.com <javascript:>.
> Visit this group at https://groups.google.com/group/cp2k.
> For more options, visit https://groups.google.com/d/optout.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20181004/6dffacf7/attachment.htm>