[gmx-users] Shifting in Verlet cut-off schemes?

[Mark Abraham]

On Sat, Mar 2, 2013 at 6:40 PM, Yun Shi <yunshi09 at gmail.com> wrote:

> Hi all,
>
> I have read http://www.gromacs.org/Documentation/Cut-off_schemes, but
> still unsure about how Verlet works.
>
> "The group cut-off scheme can be combined with a buffered
> pair-list......The Verlet list scheme has buffered neighborlists with
> exact cut-off's". What does "buffered neighborlists" mean? Including
> neighbors beyond the cutoff distance (e.g. 1.0 nm)? Then what is the
> delta-distance for this buffer region?
>

See manual 7.3.9

"LJ and Coulomb potential are by default shifted to zero by
> subtracting the value at the cut-off". Does the program only do such
> subtractions at exactly the cutoff distance? Or does it shift both
> potential functions throughout the entire distance range (e.g. from 0
> to 1.0 nm)?
>

The latter. Doing it at the cut-off would serve no purpose at all (since it
is zero after the cut-off, by definition).


> Then would this cause potential problems since some force fields are
> not parametrized with a shifting approach?
>

In principle, yes. Many force fields were parameterized using methods that
are either now totally unused, or that remain in favour pretty much because
of those force fields :-) As always in computational chemistry, you need to
seek a method reasonably likely to lead to an accurate model of reality.
Combing the literature for the necessary insight is difficult, but unless
someone has already done that work for a system similar to yours, you might
as well be throwing darts at a board as not do it! There will be no
literature for using the GROMACS 4.6 Verlet kernels, because they're new.
So there's something of a burden of proof on people adopting it before the
GROMACS authors publish their thoughts on using it. Work in progress. That
said, there's no doubting that the Verlet kernels are technically superior
to the group kernels, in theory and implementation (as discussed on the
page you linked). The case for using Verlet kernels right now is stronger
if you don't have much water, want good energy conservation, or anticipate
a long-term need for strong scaling to lots of hardware. You've no
alternative if you want to use GPUs.

Mark