[molpro-user] convergence of mcscf gradient calculation (fwd)

[Dr Seth Olsen]

Hi Rika,

I've seen this sort of behavior from time to time in my calculations as
well.  I've a (not very well developed) feeling that this sort of
pathology might arise frequently in state-averaged CAS calculations.  In
the context of molpro, I have conversations with colleagues and the
part-consensus seems to be that perhaps the initial guess for the
solution of the CP-MCSCF equations in MolPro could be improved.  This
seems to be backed up by the manual, which describes the CP-MCSCF
solution implementation as 'preliminary and not very efficient'.  

Oded might consider using an alternative software package to perform
MR-CI calculations with a reduced active space.  Depending on what
states he is investigating, etc. and what the primary CI contributions
to them are, he might have better luck without sacrificing accuracy.  At
present I believe COLUMBUS implements analytic derivatives for MR-CI
wavefunctions, though there may be others.  I believe there is a basis
set limit of about 250 functions for which it has been tested, which
doesn't suit me but may suit Oded.  Other than that, the only fix I can
think of (besides rewriting the CPMCSCF) is easing up on the convergence.  

The quality of the CASSCF solution might also change things, in which
case taking some of the steps listed at the end of the manual chapter on
the MCSCF program might be worth a shot.  

Cheers,

Seth

ccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccms

Dr Seth Olsen, PhD
Postdoctoral Fellow, Biomolecular Modeling Group
Centre for Computational Molecular Science
Chemistry Building,
The University of Queensland
Qld 4072, Brisbane, Australia

tel (617) 33653732
fax (617) 33654623
email: s.olsen1 at uq.edu.au
Web: www.ccms.uq.edu.au 

ccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccmsccms




----- Original Message -----
From: Rika Kobayashi <rxk900 at anusf.anu.edu.au>
Date: Tuesday, November 29, 2005 9:22 am
Subject: [molpro-user] convergence of mcscf gradient calculation (fwd)
> Query to the National Facility Help Desk (outputs attached).
> -----------------------------------------------------------
> We believed it was a problem to do with near-degeneracy effects
> hence the comments below. His original question was:
> "I'm trying to calculate the energy gradient and derivative 
> coupling  
> for the first 4 states of the H3 system using the Molpro package.
> i'm doing this calculation using MCSCF and the 6-311g** basis-set  
> (because  for the gradient calculation i can't use contracted basis-
> 
> set in molpro)
> however there are a few things that i don't understand....
> 
> the first one is that depending on my start guess i get different  
> results for the energies they differ by about 0.0001 hartrees
> 
> the second thing is that when i run the calculation with increasing 
> 
> number of orbitals in the active space i find that the energies 
> seem  
> to converge but the derivative and the derivative coupling don't."
> 
> 
> Forwarded message:
> > From: Oded Godsi <godsi at rsc.anu.edu.au>
> > Subject: Re: convergence of mcscf gradient calculation
> > Date: Tue, 29 Nov 2005 09:47:42 +1100
> > 
> > The point is that although the energy converge pretty good as i  
> > increase the active space the gradients and the derivative 
> coupling  
> > don't.
> > 
> > i have this other program which calculate the potential energy  
> > surface and in a way i can tell if the derivatives i get are 
> right...> so for 9, 15 and 18 orbitals in the active space i get 
> almost the  
> > right results while in all other size of active space i get rubbish.
> > unfortunately for the reaction i'm interested in i have make 
> quite a  
> > few calculation near the degeneracy geometry so is there a way to 
> get  
> > around the problems of convergence around this area?
> > 
> > thanks again
> > oded
> > 
>