[CP2K-user] [CP2K:13330] GEO_OPT a molecular is 10000 time longer then gaussian

[Thomas Kühne]

Dear Linfeng, 

a couple of comments. Setting EPS_SCF to ridiculous low values below 1E-8 makes no sense - you are basically „optimizing“ noise. 
Using OT, for a geo_opt EPS_SCF 1E-7 should be fine but you should also set EPS_DEFAULT to 1E-12. Also setting all MAX_* and 
RMS_* values to 1E-4 is a rather stringent criterion, I typically go with 1/3 of the default values, i.e. 

MAX_DR 1E-3
MAX_FORCE 1.5E-4
RMS_DR 5E-4
RMS_FORCE 1E-4

Lastly, within CP2K LBFGS is typically faster converging than BFGS, though slightly less intuitive in its usage. 

Cheers, 
Thomas

> Am 18.05.2020 um 09:45 schrieb Linfeng Gan <linfe... at gmail.com>:
> 
> Here is the result.
> 
> 
> On Monday, May 18, 2020 at 8:41:45 AM UTC+8, Linfeng Gan wrote:
> 
>  
>  
> Thank you for your advice, Travis.
> 
> I had tested EPS_SCF from 1e-5 to 1e-14. I posted the 1e-5 one only because it took the shortest time to get the final structure. And I will try convergence criteria with 1e-4 as you suggest.
> 
>  
> Linfeng
> 
> 
> 
> On Sunday, May 17, 2020 at 12:36:33 PM UTC+8, Travis wrote:
> Hi,
> 
> I see you have EPS_SCF commented out in your input file. The default convergence threshold is 1e-5 au which is usually not tight enough for accurate forces (and certainly not for vibrational analysis, which is computed numerically - hence it will take ages over Gaussian). The value you commented out is actually better and may take less steps to converge. Your criteria for convergence are also extremely tight, 1e-4 au is typically enough. Also, just to be clear, CP2K when setup appropriately will reproduce Gaussian output for the total energies for non-periodic calculations. So you CAN use Gaussian for the gas phase reaction and CP2K for the surface reaction.
> 
> -T
> 
> 
> On Saturday, May 16, 2020 at 10:47:56 PM UTC-4, Linfeng Gan wrote:
> 
> Thank you very much Vladimir,
> 
> 1) I know CP2K is not a choice for small molecules. But I am investigating an organic reaction which could take place in gas phase and on C3N4 surface. I have to compare the data between the gas phase and the surface. So, I can’t use gaussian or orca to that. DMol3 or CASTEP maybe is best choice for me, but I think we should give open source software a chance. So, CP2K is my choice right now. I am not very familiar with CP2K that it’s why I go to this forum and ask for some help.
> 
> 2) Agree. 
> 
> 3) Yes, it was. 
> 
> 4) 10000 times is an exaggeration, I apologize for that. Form the same initial structure, it took gaussian 41 seconds with a single Xeon 2680v2 CPU to finish the geometry optimization, but it took CP2k 40 minutes with 20 CPU to do the same thing. If we include the frequencies calculation the time is 4 minutes vs 71 minutesX20CPU. A single SCF calculation of CP2K was pretty fast, but most of time was wasting in locating the minimize. 
> This efficiency issue only happened in this 1-butene molecular. Another molecular I tested, cyclobutene, seemed much batter than 1-butene. 
> 
> I attach all the input and the output file. So, you guy can help me to improve this. Thanks all.
> 
> 
> Best,
> LInfeng
> 
> 
> On Friday, May 15, 2020 at 7:41:12 PM UTC+8, Vladimir Rybkin wrote:
> Dear Linfeng,
> 
> a few general remarks:
> 1) CP2K is not the most efficient tool for small molecules. For instance, it does a lot of integrations over cells in real and reciprocal space, and for a small molecule the cell is almost everywhere empty.
> 2) CP2K does not use internal coordinates for optimisation. This has to do with the fact that it's mostly meant for large periodic systems. Gaussians' Berny algorithm does use internal coordinates and is very efficient for small systems.
> 3) It's very likely that you 150 steps have not change the geometry significantly. 
> 4) It's not completely clear where this 10000 times come from. It may have to do with how you compile and run your applications.
> 
> That said, there's no specific problems with optimisers and efficiency. It's about applicability of tools for specific purposes. If you going to work with small molecules, you'd better stick to programs made for them, i.e. Gaussian. CP2K is meant for large and/or periodic systems.
> 
> Yours,
> 
> Vladimir
> 
> пятница, 15 мая 2020 г., 2:51:51 UTC+2 пользователь Linfeng Gan написал:
> Hi all,
> 
> When I optimized 1-butene molecular at B3lYP/6-311G** level, it took much much longer than gaussian16 did.  I had test optimizer CG, BDGS and LBFGD, nothing better. It seems the optimizers have some efficiency issues. All three optimizer had taken at least 150 optimization steps to get the final structure, but the initial structure had been optimized by gaussian.   How could I improve that?
> 
>  
> Best
> 
> Linfeng
> 
> 
> -- 
> 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 <mailto:cp... at googlegroups.com>.
> To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/e60c3ce3-6df6-48f1-8026-460bfc1a7798%40googlegroups.com <https://groups.google.com/d/msgid/cp2k/e60c3ce3-6df6-48f1-8026-460bfc1a7798%40googlegroups.com?utm_medium=email&utm_source=footer>.
> <opt_geo.out><opt_geo.inp>



==============================
Thomas D. Kühne
Dynamics of Condensed Matter
Chair of Theoretical Chemistry
University of Paderborn
Warburger Str. 100
D-33098 Paderborn
Germany
tdku... at mail.upb.de
+49/(0)5251/60-5726

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20200518/a9bf5cf3/attachment.htm>