[gmx-users] Energy conservation issue

Yujie Wu yujie.wu at hec.utah.edu
Fri Oct 28 22:36:05 CEST 2005


> Hi,
> 
> >
> > With a plain cut-off one of the biggest sources of error is the
> > charge group based neighborlist.

Why not remove this big source of error (BSOE), especially when
Ewald/PME/PPPM is used. (Or maybe add a mdp option for the user to
choose, but I doubt anyone would like to choose the BSOE :-) I imagine
this change just requires a few line of code.


> > When I change the water topology from one charge group to
> > three (one for each atom) or when I used a shift function for the LJ
> > I get 0 drift in double precision (300 ps) and about -85 kJ/mol/ns
> > for single precision.
> >
> > One can not expect that single precision does as good as double
> > precision. But the -85 kJ/mol/ns is already very good and comparable
> > to other errors introduced by the simulation setup, such a charge  
> > group
> > based neighborlists and PME approximations.

Thank you very much, Berk, for the check. It really helps me to
understand the issue.  If it is all due to single precision, then the
energy drifting doesn't bother me too much.

> 
> The main problem in single precision are the constraints - in testing  
> we've actually achieved perfect conservation even with constraints,  
> as long as the integration/constraint part is perfomed in double  
> (requires hacking).

I love to hear this. Many people still consider achieving (nearly)
perfect energy conservation (for an isolated system) in single precision
will be a big advantage. And I believe performing the integration and/or
constraints in double precision will have negligible overhead, so I
would love to see this "hacking" can be made default inside Gromacs.

> 
> You can also play with ewald_rtol to improve energy conservation with  
> PME. The current PME interacition form is optimized to be as close as  
> possible to the "true" infinite coulomb sum - you could envision  
> adding a small switching term to improve the energy conservation and/ 
> or zeroing the net force sum from the lattice part, but this again  
> requires hacking.

Thanks, Erik, for all these insights.

Best regards,

Yujie

> 
> Cheers,
> 
> Erik
> -----------------------------------------------------------
> Erik Lindahl  <lindahl at sbc.su.se>
> Assistant Professor, Stockholm Bioinformatics Center
> Stockholm University, SE 106 91 Stockholm
> Phone: +46 8 5537 8564     Fax: +46 8 5537 8214
> 
> 
> 
> 





More information about the gromacs.org_gmx-users mailing list