[gmx-developers] triplet range cut-off

Erik Lindahl lindahl at cbr.su.se
Fri Sep 7 13:00:40 CEST 2007


On Sep 6, 2007, at 11:33 PM, David van der Spoel wrote:
>>> Added to wiki project list... weirdly I had thought it was  
>>> already there...
>> The PME-van der waals would be fantastic. However, I understood  
>> that it would be so slow that it is basically not worth doing. I  
>> have included posts here in the hopes that somebody can figure out  
>> a better way to do it. I believe that "Particle Mesh LJ?" would  
>> really be a substantial benifit for lipid bilayers or detergent  
>> micelles. Also probably very important for proteins in general.
> It would be slow, yes, but the question is how slow? It's probably  
> good to consider how important it really is, by quantifying the  
> dispersion forces, for instances in a membrane as a function of Z- 
> coordinate by varying cut-offs. This can be done without great  
> trouble using a very a short mdrun -rerun of an existing trajectory.
> I'm pretty confident that the error due to leaving out Van der  
> Waals beyond the cut-off, in particular with the dispersion  
> correction turned on, is not very large. The obvious problem is the  
> inhomogeneity of the system, this can be quantified even easier by  
> summing the C6 terms for all atoms over slabs parallel to the  
> membrane, i.e. making a C6 profile rather than a density profile.

That thought has actually crossed my mind to - doing anisotropic  
dispersion corrections.

I won't rule out implementing LJ-PME, and I guess there are a couple  
of optimizations we could do, e.g. optionally assume geometric  
combination rules for the reciprocal space part. At least in theory  
it could also make it possible to use very short direct-space cutoffs  
(0.5nm), but that doesn't go too well with the new domain  
decomposition at the moment.

However, before embarking on it I agree we need to see a couple of  
examples that quantify the improvement you get from long LJ :-)




More information about the gromacs.org_gmx-developers mailing list