CCL:G: [SPAM] CCL:G: RI DFT approximation in G03

 Sent to CCL by: frisch[*]gaussian.com (Michael Frisch)
 On Wed, Mar 08, 2006 at 09:07:06AM -0500, Stefan Grimme grimmes..uni-muenster.de
 wrote:
 > Sent to CCL by: Stefan Grimme [grimmes---uni-muenster.de]
 >
 > Dear Pablo,
 > for acetone PBE/SV(P) I get a RI-non-RI difference of
 > 0.6 mEh with TURBOMOLE (old Coulomb aux-basis).
 > Some time ago I did a RI-test with G03 that also produced a very, very big
 > difference of some hundreds of mEh, so to me that sounds like a program
 bug.
 The way the coefficients for contracted fitting sets are specified in
 G03 is rather eccentric and accounts for this problem.  You have to
 provide the coeficients of the unnormalized primitives, which is
 cumbersome.
 The easiest way to use density fitting in G03 is to use the option to
 automatically generate the fitting set from the AO basis, e.g.
 BLYP/6-31g*/Auto
 or
 HCTH/TZVP/Auto if you prefer the Alrichs basis sets.
 The automatically generated set gives similar accuracy to the Alrichs
 fittings sets for energies.  Geometries and frequencies are virtually
 identical to the non-fitted results -- the differences are much, much
 smaller than the errors inherent in the method.  So if you're doing a
 study using a pure DFT functional in G03, you should routinely use
 /Auto rather than doing the Coulomb problem without fitting.
 Fitting also speeds up TD-DFT with pure functionals.  TD-DFT results are
 a lot better with hybrid functionals, but for medium to large molecules
 it's often faster to do the pure calculation first as a guess for the
 hybrid, e.g.
 BLYP/Aug-CC-pVTZ/Auto TD
 followed by
 B3LYP/Aug-CC-pVTZ TD=Read Geom=Check Guess=read
 Mike Frisch