[Towhee-users] Towhee 6.0.5 - chemical potential changes

[Marcus M. <mar...@us...>]

Hello Towhee Users,

In the course of implementing rigid torsions for use with the
configurational-bias Monte Carlo (CBMC) algorithms in Towhee I have
resolved an unexpected bug, and also changed a substantial number of
things that will have an affect on the results of your simulations.

First, the bug.  Turns out there was a mistake in the logic when
adding the last atom to complete a ring.  The dihedral trials were not
generated properly in some cases.  At the very least this mistake
hurts the acceptance rate of cyclic moves, and it is possible that it
caused some subtle errors in the results.  I strongly suggest
upgrading to the new version if you are simulating cyclic molecules.

Second, chemical potential.  For those who don't already know - I hate
chemical potential.  As a conceptual device it is extremely powerful
and rather important to thermodynamics so I understand the allure of
computing the chemical potential.  However, from a computational point
of view it really is a problem as soon as you start using it on any
system more complicated than single-atom Lennard-Jonesium (and even
there you can get into trouble at high densities).  The question of
how to properly normalize the bond length generations remains open
(and Towhee is currently continuing to ignore the volume-dependent
constant, but dividing out by the r^2 distribution as explained in the
Martin and Frischknecht 2006 paper).  I am now fairly certain that the
chemical potential of any rigid bond, angle or dihedral is essentially
infinite.  Alas, entropy does not look kindly on choosing a finite
number of possibilities from a continuous distribution.  Anyway, that
basically means the total chemical potential of any molecule that has
rigid interactions is nonsense.  However, if you ignore those infinite
contributions from the rigid portions you can still measure a chemical
potential and use it if you wish.  I don't advise it, but I won't stop
you either (although I sometimes may make snooty comments when I
review your papers).  On the bright side, I finally do understand how
to handle rigid interactions properly within the arbitrary trial
distribution framework.  Previously the ratio of true/arbitrary trial
distributions had been set to 1.0 when generating rigid bonds, angles
or dihedrals, but now I understand what that ratio really should be
and am using that value.  For those who had previously computed the
chemical potential of molecules with rigid bonds or angles and were
using that in Grand Canonical simulations (I am talking to you, people
who simulate TIP3P, TIP4P, and other water models) the new chemical
potentials will be different and using your old values of the chemical
potential in new simulations will result in significant deviations
from the answers you desire.  In most cases, the chemical potential
has shifted by some constant.  Please recompute the total chemical
potential before using it in a Grand Canonical simulation.

Third, rigid dihedrals are now implemented and appear to be working
very well.  The TraPPE-UA forcefield file was modified to use the
multiple rigid dihedral potential for their aromatic torsions and the
TraPPE_Isomers example now contains CBMC generated structures for
benzene, propylbenzene and napthalene.  This should make things easier
for folks working with rigid aromatic rings as Towhee can now generate
the initial structures using 'full cbmc' instead of the template
method, and this also opens up the use of the Aggregation-volume bias
moves for those molecules.

Forth, in an effort to improve the speed of the code, random numbers
are no longer generated in cases where there is only 1 trial to choose
from.  This happens a lot more than I expected as it comes up in the
dihedral selection portion of the second step of every polyatomic CBMC
insertion attempt.  This means the random number sequence is going to
be different for almost all users of CBMC.  Has no effect on the
quality of the results, but means you will not get EXACTLY the same
simulation with the new version as you did with previous versions.
Not a problem unless you have a test suite that you are checking for
reproducing the results from previous versions.  That most likely only
affects developers, but you have been warned just in case.

Finally, I have grouped together some of the essays about the
algorithms from obscure places in the Manual (previously they were
scattered around in the code manual and in the standard output manual)
and put them in a section that is accessed from the main page.  I am
hoping to further expand this section of the manual in the future and
make it a mini-review article about the primary algorithms in the
code.  Feedback on the utility of those essays is welcomed.

Marcus

-- 
Marcus G. Martin
Director, Useful Bias Incorporated
88 Martinez Road
Edgewood NM 87015-8222
ph. (505) 286-4457
www.usefulbias.com
www.photobirder.com