Re: Constraint failure in RATTLE algorithm - shouldn't a rerun fail at the same step?

From: Bjoern Olausson (namdlist_at_googlemail.com)
Date: Thu Sep 10 2009 - 05:38:18 CDT

On Wednesday 09 September 2009 16:51:42 you wrote:
> On Tue, 2009-09-08 at 10:12 +0200, Bjoern Olausson wrote:
> > Hi everyone,
>
> hi bjoern,
>
> > it is some kind of strange to me what's going on here.
> >
> > I have a membrane NPGT simulations which crashes irregularly.
> >
> > The simulation was running fine for Xns (divided into 1ns runs) the Nth
> > run crashed with "ERROR: Constraint failure in RATTLE algorithm for atom
> > 6091!" at Step 16193900. Okay, no big deal so far. I run the Nth run
> > again (from the restart files of the previous successful run). It crashed
> > again with "ERROR: Constraint failure in RATTLE algorithm for atom
> > 39616!" at Step 16337300. Now it made it over the previous crash point
> > and failed on another timestep and another atom. Rerunning the Nth run a
> > third time (again from the restart files of the previous successful run)
> > it successfully completed.
> >
> > Now my question:
> > Shouldn't the Simulation crash in all three runs at the same position an
> > atom?
>
> not necessarily. please keep in mind, that you are running with a
> langevin thermostat that introduces some randomnes. if you are running
> in parallel, you also have the load balancing that will redistribute
> atoms sometimes differently. in general, MD trajectories tend to diverge
> exponentially.
>
mmh, really exponentially?

> the real reason for the crashes have to be related, between the
> individual runs. have you tried to look up what kind of atom
> (most likely in a water molecule) is having the trouble. if you
> get shake/rattle constraint failures, that is usually a sign for too
> large forces and resulting too large displacements/acceleration
> for shake/rattle to converge. that in turn can have three reasons,
> your system is far from equilibrium, your parametrization is bad
> somewhere, or your time step is too large for a system with such
> steep potentials.
>

Your first guess could fit. I am playing around with the SurfaceTensionTarget
and it takes rather long to reach the equilibrium state for a 90x90A Membrane.

I am simulating now for 36ns and still the order parameters of my system keep
falling... sadly below my experimental ones... so I have to change the
SurfaceTensionTarget again :-(

Is there some rule or something with which I could calculate the target
SurfaceTensionTarget value when I know, that -4, 1 and 6 are dancing around at
S 0.52 (for now I can't really tell if there is a noticable difference between
those three values). My target S would be 0.232.

Thanks for the clarification
Greetings
Bjoern


This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 05:22:22 CST