[gmx-users] TI, sampling, sc_power, and sc_alpha

[Berk Hess]

>From: mernst at tricity.wsu.edu
>Reply-To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>To: gmx-users at gromacs.org
>Subject: [gmx-users] TI, sampling, sc_power, and sc_alpha
>Date: Fri, 26 May 2006 12:36:20 -0700 (PDT)
>
>Greetings once again fellow Gromacs-users,
>
>I have seen a substantial and disturbing change in calculated free energy 
>differences
>between systems run with Gromacs 3.3 and Gromacs 3.3.1. I am trying to 
>determine if
>these differences are due to the new software, new parameters, insufficient 
>sampling, or
>something else, and I don't have enough computer time to investigate every 
>possible
>combination. I turn to the list for insight.

I had implemented the alpha=1 option in 3.3 and set it as default,
since it seemed to produce "better" dH/dl curves, especially for LJ-only
perturbations.
Only some time after the 3.3 release I realized that this might cause a lot
of trouble as people probably did not realize that the default alpha value
had changed. I apologize for having caused such inconveniences.
To prevent such mistakes I have set default value to zero in 3.3.1,
which is not allowed, so people are forced to think about alpha.

The problem is that there is not (yet) an optimal soft-core dependence
for combined charge/LJ perturbations. For LJ only alpha=1 is quite optimal.
But for charge perturbations alpha=1 causes an extremely sharp peak
exactly at lambda=0 (or 1). For alpha=2 this peak at zero shift to values 
slighlty
above zero and becomes less sharp, here you have to sample the both slopes
of the peak.
With alpha=1 you need a very non-linear lambda distribution with a very
fine spacing close to 0. With alpha=2 you also need a fine spacing at 0,
but probably also in the middle where the particles start overlapping.

I would guess that especially alpha=1 can give an enourmous error
if you only used 21 evenly spaced points. Adding points at 0.01 and 0.02
from lambda=0 and/or 1 could help a lot.

Berk.