> I have a question about restarting the simulation from the same restart
> file when using NPT with Langevin dynamics (and Berendsen barostat).
> Assuming that each run with Langevin dynamics is essentially different
> (or at least I believe so if the ig value is different), should the
> trajectories obtained from the same restart file (with irest=1) be
> actually different or are they expected to be "the same" (within
> numerical precision errors etc.)? By "the same" I mean e.g. from the
> point of RMSD. I ran 3 replicas restarting from the same restart file,
> and even though they behaved similarly from the beginning, after some
> time, they started "diverging". The changes are smaller when comparing
> it to a new run with reinitialized velocities but still. Is it normal?
> Can anyone explain what to expect and what is not OK anymore (if it can
> be said)?
Langevin dynamics relies on the random seed, and if different, the
trajectories will diverge (i.e. not be the same). Berendsen does not rely
on a random number so the trajectories will not be expected to diverge
rapidly, however if running on a CPU due to non-deterministic nature of
floating point operations, the trajectories will in time diverge. If
running on a (single) GPU with the mixed single / fixed precision model
(default) the Berendsen trajectories will not diverge as that code is
deterministic (and therefore a great way to test for GPU failures or
errors). Same is also true on the Anton MD engine which is why if you run
on Anton, you do not dump the trajectory too frequently since it slows the
code down. If you find some rare conformational transition you want to
probe deeper, you go back to the restart file before and re-run dumping
the trajectory at a finer grain.
--tec3
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Dec 03 2023 - 13:30:02 PST
