[gmx-users] Electrostatic forces

[Mark Abraham]

Berk Hess wrote:
Hi Berk,

Thanks for your feedback.

> I think there is still a general issue with optimizing the PME settings
> (or settings for any electrostatics method) for MD simulations.
> The question is what exactly one should optimize.
> If you do a single point calculation, you might only want to minimize
> the force error. But for MD you might also want good energy conservation.
> Optimizing the RMS force is certainly not the best choice,
> since that would lead to a large ewald_rtol, which gives small
> errors before the cut-off and on most of the long-range part,
> but large errors just beyond the cut-off. This will lead to bad
> energy conservation. Even an absolute force error will probably
> lead to a too large ewald_rtol.

I think the optimization target function should be the RMS error over 
the 3 components of the total force on each of the N atoms. Thus the RMS 
is over 3N values. I can't see how this optimization process would seek 
an (rcoulomb, ewald_rtol) point where large errors at r = rcoulomb+delta 
exist, since this large error would propagate immediately to the error 
in the total force on each atom, and persist through the subsequent RMS 
averaging, since all atoms are affected similarly. If the RMS averaging 
was over all contributions to the components of the total force on each 
atom (which is not well-formed for Ewald anyway) then I could maybe see 
such a problem arising.

I do agree that such an optimization should check for suitable energy 
conservation also. Someone planning NVE has different requirements from 
NVT, etc, and so they may need to optimize differently.

Mark