I'm looking to get your opinions on the most promising density functionals to use for thermochemistry and kinetics of transition metal complexes. However, as eloquently laid out by Tom Manz on this CCL post, I am not solely interested in the best functional according to a given benchmark set. As a practical user, there are several other important considerations I'd like to emphasize given my needs:
There must be support for geometry optimizations and analytical vibrational frequencies for this functional in either ORCA or Gaussian, the former because it is free and publicly available and the latter as a last resort because many computing facilities have a license. For ORCA, this requirement sadly rules out ωB97X-V and ωB97M-V (no support for gradients with nonlocal correlation) and all meta-(hybrid)-GGAs (no support for analytical frequencies). Naturally, this requirement also implies that the functional must be present in ORCA (including via libxc) or Gaussian as well. For Gaussian, meta-(hybrid)-GGAs have support for analytical frequencies, but there is a much more limited selection of functionals (e.g. ωB97X-V and ωB97M-V are not present).
Converging the self-consistent field (SCF) should be as minimal a nightmare as possible. Older meta-(hybrid)-GGAs, such as M06, can be particularly problematic here. This is largely resolved with newer Minnesota functionals that include smoothness constraints, such as revM06 and MN15, but these functionals violate Condition #1 and #4, respectively.
Ideally, I'd like it to also be reasonable at capturing spin states of transition metal complexes. This is perhaps asking for a lot, but I put it down anyway. M06-2X, for instance, is shown to be good at predicting kinetics for organometallic reactions, but the very high fraction of HF exchange (i.e. 54%) is concerning for other properties.
If given the option, I'd prefer a functional that does not have an exorbitant number of fitted parameters. I will always have some degree of hesitation with MN15, for instance.
On one hand, a Minnesota functional would seem to be ideal based solely on the title of this post alone. However, the fact that all of them are meta-(hybrid) functionals rules out their practical use with ORCA, and the newer revised versions of the M06 family are not present in Gaussian. On the other hand, I like what I've seen with the range-separated hybrids ωB97X-V and ωB97M-V, but both can't readily be used in ORCA or Gaussian as mentioned above.
I feel like this mostly leaves me with ωB97X-D as the clearest choice to consider. However, I guess this would be sacrificing the greater accuracy of ωB97X-V for the availability of ωB97X-D. In terms of modifications to ωB97X-D, to my surprise ωB97X with D3(BJ) corrections does worse than the standard -D (i.e. D2) correction for barrier heights according to this paper. Will need to investigate this more.
Any suggestions are welcome.