Re: Convergence of ABF calculation: n_samples

From: Chris Chipot (
Date: Wed Dec 05 2007 - 09:53:39 CST

Dear Vaithee,

by and large, we advocate to break down the reaction pathway into
contiguous, *non-overlapping* windows when the former spans a wide
range of xi-values. The purpose of this is essentially to improve
the efficiency of the free energy calculation, with the hope that
the time required for all simulations to reach a uniform sampling
in every individual windows be shorter than that needed in a long,
unique simulation covering the entire reaction pathway.

There is a caveat, however, that was shown by Jerome and I some
time ago. Quasi non-ergodicity effects can be observed in such a
stratification strategy, should *xi* be strongly coupled to other
slowly relaxing degrees of freedom. A naive example consists of
parallel narrow valleys along *xi* separated by high free energy
barriers. While ABF enhances diffusion along *xi*, sampling in the
other dimension would be strongly hampered, yielding inaccurate
estimates of the free energy derivative. Whereas these barriers
can be moderate in a given window, they can be substantial in
others, which generally results in an incomplete sampling.


Subramanian Vaitheeswaran a écrit :
> Dear Chris and Jerome,
> I have a few questions regarding the optimization of an ABF calculation. When would it be a good idea to break the reaction coordinate into multiple windows? - when the free energy changes along this coordinate are large?
> If I do have multiple windows, I can't use the uniformity of n_samples to judge the convergence of the pmf. Would it be enough if sampling is uniform (or approximately so) *within* each window?
> thanks,
> Vaithee


Chris Chipot, Ph.D.
Equipe de dynamique des assemblages membranaires
Unité mixte de recherche CNRS/UHP No 7565
Nancy Université Phone: +33 (0)3-83-68-40-97
B.P. 239 Fax: +33 (0)3-83-68-43-87
54506 Vandœuvre-lès-Nancy Cedex


     Science without management is worse than management without science

                                                        N. G. van Kampen

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:45:39 CST