[gmx-users] Reproducing results with independent runs

[Mark Abraham]

On Sat, Nov 9, 2013 at 9:36 AM, alex.bjorling <alex.bjorling at gmail.com>wrote:

> Dear users,
>
> I am investigating protein crystal packing artifacts by doing equilibrium
> simulations starting from a crystal structure. I would like to know if the
> relaxations i see are reproducible, in the sense that many simulations with
> independent velocities give the same general result.
>
> My plans is to do only one set of (first NVT then NPT) equilibrations with
> position restraints. Then, I thought I'd do a shorter NPT run with position
> restraints, with more frequent output and using the trr snapshots as
> starting points for production runs.
>
> The only question then is how far apart these snapshots need to be to
> guarantee independent velocities. Attached is the velocity autocorrelation
> for the Protein group. It seems to me that using snapshots 1ps apart would
> do it, since the autocorrelation has decayed by then.
>
> Is this a valid approach?
>

That the observations are uncorrelated (because the autocorrelation time
has elapsed) does not imply that trajectories started from successive
snapshots would be independent - in the absence of floating-point or
load-balancing artefacts leading to numerical divergence, the separate
simulations would nearly reproduce each other! Since you have to wait for a
period of divergence anyway, you might as well generate new velocities
after an initial stage of equilibration, equilibrate further, and have no
independence question to answer.

Mark