unique spline generation for LJ

Teodoro Laino teodor... at gmail.com
Sun Jun 24 10:37:48 CEST 2007


Dear all,

it is available (as an option.. defaults have not been changed) the  
possibility when running MM (or QMMM jobs) FF using lennard-jones  
potential to generate only 1 spline instead of the several thousands  
usually generated.. All other values (different sigma and epsilon)  
are computed from this global optimal spline for LJ.

The method is ~ 10-15% slower (tested on JAC) in the evaluation of  
the nonbonded, so at the moment, unless there won't be speed-up
improvements on this part of the code, the unique generation of the  
spline for LJ can be requested by the use but won't be the default.

You may experience differences in numbers (for default values)  
between this implementation and the old one.. This difference becomes  
small when tightening the eps_spline value.. For very small  
eps_spline (up to 5*10^-14.. the limit I was able reach before  
numerical noise was creating troubles) the new method converges  
quickly and with mostly no cost to the analytical LJ value. For the  
old implementation it may be a little bit more problematic, since
generating a spline with an eps of 5*10^-14 requires at least 1-2  
millions of points and as in the case of JAC the number of spline is  
2500.. so it may be quite expensive (cpu and memory) to generate the  
full set of spline for that accuracy..

I would highly suggest to all the people doing biostuff (MM or QMMM)  
to use the new option with small values of EPS_SPLINE (10^-12 should  
be safe)..
The new implementation can be switched on activating the keyword:

&SPLINE
    UNIQUE_SPLINE
&END

This does NOT apply If you don't use LJ potential but other non- 
bonded FF.
Feedback is very welcome!
Cheers
Teo

------------------------------------------------------------------------ 
----------
Teodoro Laino                                   Phone : ++41 44 635 4479
Physical Chemistry Institute              FAX   : ++41 44 635 6838
University of Zurich                             E-mail:  
tla... at pci.uzh.ch
Winterthurerstrasse 190
CH-8057 Zurich, Switzerland

CP2K Google Forum:     http://groups.google.com/group/cp2k
HomePage:                                        http:// 
www.pci.uzh.ch/laino
------------------------------------------------------------------------ 
----------







More information about the CP2K-user mailing list