[gmx-users] soft-core potential in combination with PME

[Jeroen van Bemmelen]

Hi David and all,

Tanks for your elaboration on the subject. I've done a couple of free 
energy simulations myself, also on disappearing charged atoms, but so 
far I've not yet encountered any of the instability problems you 
mention here (though I did encounter instabilities at high lamda 
values, caused by the flying ice cube problem). Your problem could 
very well be related to the force field, since I'm using G53a6 and 
you're probably using OPLS. Maybe the extra hydrogens are causing 
your troubles?

I agree with you that the ideal soft-core parameters for decoupling 
LJ may not be so ideal for decoupling Coulomb, and that this can lead 
to convergence problems and wrong free energy results. In that case a 
separate decoupling of LJ and Coulomb forces in fact may be the best 
answer.

Actually at first glance I found the Anwar paper to be quite 
interesting in this respect, because it uses a different functional 
form for damping the short-range Coulomb force when using PME, that 
specifically 'fits' the Ewald real-space term. And indeed it gives 
you an additional parameter (aqq) for separately tuning the Coulomb 
soft-core potential. But since the standard GROMACS soft-core should 
also remove the singularity and since I haven't had any convergence 
issues up till now (or maybe I'm just simulating too long ;-) ), I 
guess I don't need Anwar's method.

Btw, David, are you still interested in feedback on your free energy 
tutorial, or do you consider that project finished? I used it as a 
starting point for my simulations and may have some interesting 
remarks and/or suggestions.

Greetings,
Jeroen


> Date: Fri, 13 Apr 2007 09:18:21 -0700
> From: "David Mobley" <dmobley at gmail.com>
> Subject: Re: [gmx-users] soft-core potential in combination with PME
> 	(sorry,	again)
> To: "Discussion list for GROMACS users" <gmx-users at gromacs.org>
> Message-ID:
> 	<bc2c99750704130918j7e8add8ene9df87cb93bb19be at mail.gmail.com>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
> 
> Berk and all,
> 
> > I don't understand what David Mobley meant exactly.
> > There is no Coulomb singularity with soft-core.
> > Maybe one could have an unfortunate situation where the LJ
> > is already very soft, but the Coulomb not very soft,
> > which could lead to instabilities.
> > But I have never encountered this.
> 
> Anytime I do a simulation where I turn of both coulomb and LJ
> interactions at the same time, I run into this problem. Sometimes it's
> worse than others. Maybe it's because the 1/r^12 and 1/r have
> different r-dependencies? I don't know.
> 
> I suspect the issue partly is just what's optimal: The soft core
> parameters that would be optimal for modifying Coulomb interactions
> are not optimal for modifying LJ interactions, and vise versa. So if
> you use soft core settings that give you a smooth transformation for
> LJ interactions, you do too  much or too little smoothing of the
> Coulomb interactions and introduce large Coulombic forces. On the
> other hand, if you pick soft core parameters that are good for Coulomb
> interactions, you end up with large LJ forces. (For example, Coulomb
> transformations are nearly optimal with LINEAR lambda scaling
> (sc-alpha=0), but that doesn't work *at all* for LJ transformations.
> 
> If I remember correctly, basically what the Anwar paper tries to
> achieve is separately smoothing the two. You could probably accomplish
> the same thing by allowing separate sc-alpha and sc-power for Coulomb
> and LJ interactions so they can be tuned separately.
> 
> 
> So, while formally using soft core for Coulomb removes the
> singularity, in practice I think the forces are still large enough
> (when using soft core parameters are tuned for LJ interactions) that
> instabilities result (at least for the stuff I do, with the force
> field I use). Hence my recommendation to do two stages.
> 
> Anecdotally, I should mention that a bunch of different people have
> e-mailed asking me about various problems they're having where they
> get unexpected free energies, etc. I suggest they do the Coulombic and
> LJ parts separately, and invariably they write back that it works much
> better.
> 
> David