Re: ABF simulation with extendedLagrangian on keeps crashing

From: Giacomo Fiorin (giacomo.fiorin_at_gmail.com)
Date: Thu May 05 2011 - 16:22:09 CDT

Hi Asjasja, just a check: are the dummy atom and the other atoms 100 Å away
even considering periodic boundary conditions? The policy within the
colvars module is to use PBCs to compute distances, as in all of the
physical interactions (vdw, Coulomb, ...).

Some times you may want to not make your colvars PBC-aware, because you
consider groups of atoms that altogether span more than half the box size.
In distance-based components, we introduced the "forceNoPBC" option, but
"angle" and "dihedral" unfortunately do not have (yet) such option because
most of their usage is for bonded systems whose structure keeps them away
from singularities. If PBCs were the issue in your case, Jérôme or I can
send you a patch to temporarily introduce the "forceNoPBC" option for
"angle" and "dihedral".

Regarding the theta = 0° and theta = 180° singularities, a division by
exactly zero practically never occurs. Only a division by a very small
number, which is perfectly legal, arithmetically speaking: it just doesn't
play nice with the finite integration time step. Since your system doesn't
naturally stay away from the singularities, as it happens in most cases, I
would definitely go for setting boundaries that do that, like you said.

Giacomo

---- ----
  Giacomo Fiorin - Postdoctoral Researcher
  ICMS - Institute for Computational Molecular Science - Temple University
  1900 N 12 th Street, Philadelphia, PA 19122
  giacomo.fiorin_at_temple.edu
---- ----

On Thu, May 5, 2011 at 4:50 PM, Ajasja Ljubetič
<ajasja.ljubetic_at_gmail.com>wrote:

> Hi Jerome,
>
> As promised, I have just implemented ABF-specific routines for the
>> angle component. I've only had time for very basic testing though. It
>> looks fine, but it would be best if you can try it on a trivial system
>> first. The two source files attached should be sufficient to enable
>> it.
>
>
> Thank you very much! Would you recommend that I compile this with the
> latest CVS or the 2.8b1 release?
> The most trivial system I can think of is the H-O-H angle bending of a
> single water molecule. The shape of the free energy well should be parabolic
> (as approximated by the CHARMM force field) with some small additions of the
> H-H non bonding interactions, which can be calculated analytically for this
> simple case. Would this make a good test case?
>
> Your crashes might come from the singularity that occurs if one of the
>> actual atoms gets too close with the dummy atom.
>
>
> The dummy atom is 100 A away, so the other atoms can't get too close
> (in Cartesian space).
> There are actually two singularities that I forgot to consider: at theta= 0
> and theta=180, where the phi angle is undefined (i.e. where three atoms of
> the dihedral component become collinear). I'll try to do a run with more
> restricted theta values.
> But I find it strange, since I was applying the bias only to the theta
> colvar (and calculating a 2D histograms of theta and
> phi), shouldn't NAMD complain about a division by zero, instead of blowing
> up velocities?
>
> Thank you & best regards,
> Ajasja
>

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:20:12 CST