[CP2K:1966] Re: Geometry Optimisation / Frequency (general question)

Laino Teodoro teodor... at gmail.com
Mon Apr 6 18:25:00 UTC 2009



> I remembered from your last reply that DX 0.001 is ok, so I did not  
> mess
> around with that one.

DX is strictly connected to the smallest frequency you want to  
compute and of course
EPS_SCF depends on it.. you may even try smaller DX, decreasing  
further EPS_SCF..
although at a certain point  the numerical error will just blow..

> One other question, number of replicas is the number of the CPUs  
> used? Or did
> I get that wrong?

number or replicas means that you can organize your N procs in  
NREPLICA groups and
each of them is doing an energy/force calculation in parallel. That's  
very convenient if you
have plenty of procs because you can get a frequency calculation done  
in few minutes even
on large molecules..
so.. let's say that you have 1024 procs available and for each of  
your energy/force calculation you
need 64 procs then you can setup 16 REPLICAs (64*16=1024) using:

NPROC_REP 64

and you will get automatically 16 replicas.

Cheers
Teo


>
> Thanks for the quick reply Theo!
>
> On a more personal side, I hope that nobody you know was effected  
> by the earth
> quake today in Italy!
>
> All the best!
>
> Jörg
>
> On Montag 06 April 2009 Teodoro Laino wrote:
>> Hi Jörg,
>>
>> I'm a bit confused:
>>> I have managed to get a converged geometry with these parameters:
>>> &GEO_OPT
>>>     MINIMIZER BFGS
>>>     MAX_ITER   300
>>>     MAX_DR     3.00E-4
>>>     RMS_DR     1.50E-4
>>>     MAX_FORCE  4.50E-5
>>>     RMS_FORCE  3.00E-5
>>>     &BFGS
>>>     &END
>>>   &END GEO_OPT
>>>
>>> and
>>> EPS_SCF 1.0E-8
>>
>> so you used 1.0E-8 for GEO_OPT. and then you did with the same  
>> EPS_SCF
>> the frequency analysis and you got
>> a neg. freq. then you decreased the value to 1.0E-8 (which is the  
>> same
>> as before :-) ) and it solved your issue..
>> well.. then you were just lucky and you may have had a problem  
>> with the
>> interaction of neutrinos with your CPU ;-).
>>
>> I suspect your first EPS_SCF was the default 1.0E-6. right?
>>
>> in any case (general rule) if the numerical frequencies are  
>> computed by
>> finite differences of course you need to
>> be careful, especially for low-frequencies. Infact, depending on  
>> DX and
>> EPS_SCF you may just observe numerical
>> noise for small frequencies.
>> You did it right then!
>>
>> Cheers
>> Teo
>>
>>> Using the same EPS_SCF 1.0E-8 for the frequency analysis, I get  
>>> one neg.
>>> freq. at -42.76 1/cm. However, as I remembered from the ADF  
>>> program, it
>>> might be actually beneficial to use a tighter criterion for the  
>>> numerical
>>> frequency analysis, so I decreased the value to EPS_SCF 1.0E-8  
>>> and loo!
>>> now that neg. frequency has gone.
>>>
>>> Is that generally a good idea to run the frequency calculation a bit
>>> tighter than the geometry optimisation or was I just lucky?
>>>
>>> I just try to get the grips of the program, hence the question.
>>>
>>> Thanks for your help!
>>>
>>> Jörg
>>
>>
> -- 
> *************************************************************
> Jörg Saßmannshausen
> Research Fellow
> University of Strathclyde
> Department of Pure and Applied Chemistry
> 295 Cathedral St.
> Glasgow
> G1 1XL
>
> email: jorg.sassm... at strath.ac.uk
> web: http://sassy.formativ.net
>
> Please avoid sending me Word or PowerPoint attachments.
> See http://www.gnu.org/philosophy/no-word-attachments.html
>
> >




More information about the CP2K-user mailing list