g90 error/SCF convergence

[friedman()at()tammy.harvard.edu (Dawn Friedman)]

   Thank you to the many who told me about "limit coredumpsize 0"  --
 among
 others, David Bernholdt at the QTP, Robert Jones at Thinking Machines,
 Terry Coley at Caltech, and Nico van Eikema Hommes at Erlangen.  Thanks to
 the "limit" command, I now know I won't be getting giant core dumps.
   The SCF convergence problem was less popular, and I'm still a little
 confused by the responses, so I thought I'd report them here and see if
 they can be brought into agreement.
   Margot at eorgia Tech says that UHF may not be a good wavefunction for
 this radical anion, and I may need to go to MCSCF or CASSCF (both in
 Gaussian, though I see that you need G92 for analytical CASSCF second
 derivatives for frequencies and force constants.)  Even if I manage to
 overcome the convergence problems at UHF, if those problems are chronic,
 the UHF wavefunction I obtain may be junk -- unstable or otherwise poor.
 Well, the convergence problems _are_ chronic and the neutral molecule
 wavefunction back at 3-21g* is unstable, which seems ominous.
    Nico van Eikema Hommes suggests that I may be getting an error he's
 familiar with, in which the SCF converges but then the "last cycle without
 DIIS" fails -- a problem that extended basis sets like 6-311++g** are prone
 to because of linear dependency.  The result looks like:
    SCF DONE  E(UHF)=[number]      after nn cycles
    Convergence failure, run terminated
   This is indeed what I'm seeing.  He suggests that I try the option
 IOP(5/20=2) which avoids that last cycle without DIIS.  (And notes that
 if I'm using SCF=direct, I should use SCF=(DIRECT,NOVARACC) as well.)
   "A more dangerous technique" is to force the optimization to
 continue
 despite the convergence failure, using IOP(5/13=1).  This will help if it's
 the bad geometry that's causing the convergence failure, but it may also
 "yield nonsense."
   As for me, it's clear that the first step is to figure out the instability
 of the neutral molecule wavefunction, and I've had some advice about that,
 too, which I'll be looking at.
   But can someone provide a consensus on what to do about bad SCF convergence?
 That is, if I use one of the techniques to avoid the convergence problem,
 should I then test the stability of the resulting wavefunction?
   Many thanks,
   Dawn
 ---
 Administrivia: This message is automatically appended by the mail exploder.
 CHEMISTRY()at()ccl.net --- everybody; CHEMISTRY-REQUEST()at()ccl.net ---
 coordinator only
 OSCPOST()at()ccl.net : send something from chemistry; FTP: www.ccl.net
 ---