[gmx-users] Fwd: FW: PME settings in free energy calculations

[Mark Abraham]

Hi,

On Tue, Feb 9, 2016 at 1:30 PM Dries Van Rompaey <dries.vanrompaey at gmail.com>
wrote:

> Hi,
>
> I'm not deliberately triggering the generic kernel.


OK, well that means its a setup that we have deliberately not tried to make
fast in the group scheme, because it's too hard to do for the expected
usefulness.


> I unfortunately have no
> idea what's causing that. I uploaded logfiles for a setup with matching
> cutoffs as setup3alt and setup2alt (I got quite a lot of warnings for this
> setup, which is why I initially changed the cutoffs slightly).
>
> I'm a bit baffled what could cause the large difference between the verlet
> setup with long cutoffs and the verlet setup with shorter cutoffs and PME..
> After all, that's the one that should be fairly close in order to prove my
> electrostatics treatment is appropriate. This difference seems to be a lot
> more pronounced than the difference to the group scheme.
>

That's the easy one. You're comparing potentials from Verlet PME with a
short-ranged potential shift with Verlet plain cutoffs with no potential
shift. The shift is quite small with PME (that's what ewald-rtol is all
about) but it applies to every single in-range interaction. Your
long-cutoff calculation has a cutoff that is only ~3 times longer than the
modified short-ranged form you used with PME. That increase in range is not
going to get down to 1e-5 relative error from truncation at the cutoff, and
it also applies to a lot more interactions. I expect that this is
confounding all your comparisons with Verlet+PME - you need to compare with
a scheme with a potential shift. (This also leads people astray when they
are e.g. comparing total energies between GROMACS and codes that don't use
the potential shift.)

Whether such comparisons as alchemistry.org suggests worked reasonably well
with Group+PME vs Group+long-cutoff (because neither used a potential
shift) with current or old code, I don't know.

Mark