[molpro-user] F12 methods

[Gerald Knizia]

On Sunday 19 October 2008 06:22, Christian Evenhuis wrote:
> I am trying out the new F12 methods but I am not having much luck.
>
> 1. if I use the VNZ-F12 basis sets I get the following error:
> 		1PROGRAM * CCSD (Closed-shell coupled cluster)     Authors: C.
> Hampel, H.-J. Werner, 1991, M. Deegan, P.J. Knowles, 1992
>                 CCSD-F12 implementation by  H.-J. Werner, 2007
>                 Density fitting integral evaluation by F. R. Manby,
> 2003,2007
>                 Cannot find default basis VDZ-F12 for atom O

We forgot to define default auxiliary basis sets for the VnZ-F12 orbital basis 
sets in the release version of Molpro 2008.1. I think Prof. Kirk Peterson 
fixed this with a patch in the meantime.

As a bypass, you can specify the basis sets manually like this:

   df-rhf,basis=VTZ
   explicit,df_basis=AVTZ,ri_basis=VTZ,basis_exch=VTZ
   df-mp2-f12

Typically the Weigend/JKFIT basis sets can be used both for Fock matrix 
fitting (basis_exch) and as RI basis. For the latter application new 
dedicated RI basis sets are currently being developed by Kazim Yousaf and 
Kirk Peterson.

Until these are available, it might happen regularly that functions need to be 
deleted in the CABS construction. In general this is not good and may lead to 
strange looking total energies like for the AVQZ set in your H2O-dimer 
example. Relative energies, however, usually are not much affected by this.
Anyway, you can turn on a CABS construction which is less likely to delete 
functions by

   explicit,ortho_cabs=1,thrcabs_rel=1e-9,thrcabs=1e-9

This will basically include all functions in the CABS, which are not so 
linearly dependent that they will definitely blow up numerical stability.

With the new dedicated RI sets, these problems will most likely vanish 
automatically, until then, the setting mentioned above has been found to work 
well in a great variety of different cases.

> 3.  For CCSD(T)-F12 / AVNZ  I get the following error
>               ? Warning
>               ? Encountered below-occupied eigenvalues in virtual
> block. This may indicate serious deficiencies in the orbital or
>                  JKfit-basis sets (basis_exch). Offending values will
> be ignored, but PT2 is ill-conditioned!
>               ? The problem occurs in f12_common/Mp2Singles()
>               FockProj BlockDiagonalize() eigenvalues:
> (norb=                   314 )
>               E(Occ):   -20.5969651851479       -20.5474946885548
>                -1.38499649063043       -1.33164908493971
> -0.749957814812005
>               -0.701772636068004      -0.612407775875517
> -0.558616202902750
>               -0.531801726963611      -0.483450539393589

This is not supposed to happen. Could you provide the full input for this? I 
cannot reproduce it. When I just enter "AVTZ" for BASIS and "ccsd-f12" for 
"METHOD-F12" in your template the calculation runs fine on my machine, and 
everything looks correct.

>      Changing to RI_BASIS=AVTZ/ORBITAL helped a little but did not
> resolve the issue

You should not specify orbital basis sets as RI basis set: During the 
construction of the complementary auxiliary basis set, all functions which 
can be represented in the AO space will be deleted from the RI basis (because 
the actual RI is always done in the union of the RI and the AO basis).

I'm not sure if the specification above would define an contracted or 
unconstracted aug-cc-pVTZ as RI basis (the first one would result in zero 
CABS functions), but both are definitely not good for this purpose. The 
VTZ/JKFIT which is used as default should work fine.
-- 
Gerald Knizia