Gaussian and CAS
Gaussian Users:
Steven Bachrach recently noted that the MSRC Ab Initio Benchmark Report
mentions that we obtained the wrong CASSCF energy from both Gaussian 90 and
Gaussian 92. That's true, but I may be able to shed a little more light on
it. The particular calculation was intended to exercise each program's CAS
routines. It included 4 electrons in 4 orbitals (sigma, sigma*, pi, pi*
orbitals in ethylene) and used converged RHF orbitals as the initial guess
for all programs (GAMESS-US, HONDO, MOLPRO, GAMESS-UK, etc.)
When Gaussian converged to an energy which was higher than the other codes I
assumed that it was the result of an input error. After puzzling over it for
some time, I showed it to several of the other quantum chemists here in the
MSRC. No one had any idea what was wrong. I then contacted Doug Fox via the
Gaussian Help line. However, we weren't able to obtain the correct answer.
Recently Gary Trucks from Gaussian has come up with a suggestion for an
initial guess that results in the correct energy, but at the expense of a
large number of iterations. My understanding is that they're still looking
into the problem. Perhaps someone from Gaussian, Inc. can comment here.
In Gaussian's defense, the benchmark report shows that the program performed
more of the types of calculations which people are typically interested in
doing than any of the other codes which were tested. And, while not the
fastest in every area, it turned in respectable performance on every platform
that we had, i.e. SGI, IBM, HP, Cray, DEC, Sun.
David Feller
Molecular Science Research Center
P.S. Copies of the benchmark report are available for mailing. I encourage
contributions for the next edition. There are too many machines and codes
for me to run them all.