On Tue, 10 Jul 2001, Reinaldo Pis Diez wrote:

> Dear folks,
>
> Several days ago I posted a question regarding the errors
> in frequencies calculated using the COSMO model as implemented in Adf
> 2000.02. The only answer I received is from Cory Pye, who
> implement the COSMO model into the ADF package. As a_very_short_summary I
> can say that there's no reason to expect additional "noise" introduced by
> the solvation model during a frequency calculation.
> Below is my original posting, Cory's answer follows, and a
> final comment is appended by me.
> Regards,
>
> Reinaldo
>
> > > Dear Adf'ers,
> > >
> > > I'm doing some calculations on an organic ligand (30
> > atoms) for which
> > > solvent (water) effects should be important.
> > > Geometry optimizations are ok, very cpu-time consuming
> > though. When I run
> > > frequency calculations on some conformations of the above ligand,
> > including
> > > solvent efects as well, one to three imaginary frequencies arise for all
> > > the conformations studied. The imaginary frequencies range from 30 to 110
> > > cm-1, that is they represent very floppy modes. I distorted one of the
> > > above conformations (having one imaginary freq) following the imaginary
> > > mode but I dind't succeed, a new imaginary freq arises.
> > > Now my question: normal, gas-phase, frequency
> > calculations involving
> > > numerical second derivatives are subjected to errors of +- 50 cm-1, aren't
> > > they? Does the inclusion of the COSMO model affect the error bars for
> > > frequencies? That is, is it safe to say that a 100 i cm-1 frequency is an
> > > artifact when solvent effects are introduced variationally into the KS scf
> > > equations?
> > >
>
>
>
> >Hello, I am the author of the COSMO routines in ADF.
> >The ADF COSMO routine was not tested with the FREQUENCIES option, I assume
> >you
> >mean Numerical Differentiation of the analytic 1st derivatives and not fully
> >analytic frequencies which I believe will be in some future version. There is
> >no reason a priori why the KS SCF equations should be affected to any
> >degree by
> >this, but what I would be more concerned about is the
> >creation/annihilation of
> >points on the sphere. This could cause the gradients to be a little
> >discontinuous. One of the last things I did as a post-doc with Tom Ziegler at
> >the UoC is to add a new command called FDIV=X. The default is FD=1 (60 points
> >per sphere), you might get better numbers if you tried FD=2 (240 points per
> >sphere). I don't know of these changes were incorporated in the latest ADF
> >release. This is kind of like the INTEGRATION=X (or was it ACCURACY=X?)
> >except
> >for surface points.
>
>
> Now it's me again. Adf 2000.02 has a subkey within
> SOLVATION named DIV, which controls the technical stuff to construct the
> surface. Moreover, it has an option named NDIV, which "controls how fine
> the spheres describing the surface are partitioned in small triangles, each
> containing one point charge to represent the polarization of the cavity
> surface" and has a default value of 3. I guess that these keywords are
> equivalent to FDIV and FD, the keywords mentioned by Cory.
>

Actually FDIV and NDIV are different. The way the GEPOL algorithm works is
that it initially tesselates on a sphere about each atom (60 points,
Buckyball coordinates). Then, based on the NDIV command, it further divides
these triangles by a factor of 4^(NDIV-1). Then, it determines which of these
subpoints survive (this depends on the algorithm: vdW, Klamt, SAS or SES). It
then attempts to recreate all of the original points. If all of the subpoints
associated with a point survive, then the final point is the same as the
original point. If none of the subpoints survive, the original point is
deleted. If some of the subpoints survive, the final point is the average of
all surviving subpoints. The area associated with each final point (triangle)
is the sum of the areas of the subpoints (subtriangles). This is needed in the
construction of the diagonal elements of the A-matrix.

What FDIV does is map the subpoints to another set of subpoints. It must be
less than or equal to NDIV.

Example:
NDIV=3, FDIV=2

The original 60 points (set A) are divided once to give 240 points (set B),
then divided again to give 960 points (set C). The GEPOL surface is created. In
the usual algorithm, the surviving subpoints C are mapped back to set A, but
with FDIV=2, they are mapped back to set B. The net result is that there are
approximately 4 times as many points and the gradients should be less noisy,
but more expensive to calculate. It's not elegant, but it did work in the
development version of ADF98. I don't know if this stuff was incorporated into
ADF2000. You would have to ask Steve van Gisbergen, since its been over two
years that I worked on this (July 2-July 16, 1999) and I moved to Halifax at
the end of July to take up an assistant professorship. I no longer have access
to the code.

>

   ************* ! Dr. Cory C. Pye
 ***************** ! Assistant Professor
*** ** ** ** ! Theoretical and Computational Chemistry
** * **** ! Department of Chemistry, Saint Mary's University
** * * ! 923 Robie Street, Halifax, NS B3H 3C3
** * * ! cpye@crux.stmarys.ca http://apwww.stmarys.ca/cpye
*** * * ** ! Ph: (902)-420-5654 FAX:(902)-496-8104
 ***************** !
   ************* ! Les Hartree-Focks (Apologies to Montreal Canadien Fans)
Received on 2001-07-11 16:54:38