Mixing functionals in TD-DFT opt

Question

I work with copper complexes containing at least 100 atoms (including a single Cu, nitrogens, carbons, and hydrogens). According to Grimme's paper, the benchmark level for geometry optimization is PBE0-D4/def2TZVP. However, my system is too large for this level, so I use the other option he suggested, namely, (m)GGA with a TZ basis set. The literature adviced TPSS, so I chose it.

My studies also involve a TD-DFT component, which includes, among other things, the computation of intersystem crossing rates. For this purpose, ORCA requires Hessians. The problem arises because the best reproduction of the experimental spectra was achieved with PBE0. Using PBE0 to optimize my system in the S1 and T1 states and subsequently obtain .hess files is computationally demanding for my HPC resources. Moreover, unlike Gaussian, ORCA does not allow reading excitations from a checkpoint file, so I cannot reoptimize in TPSS without recalculating excitations. This means the S1 and T1 states obtained by TPSS would be considered.

What should I do now?

Answer 1

You have run into the eternal conundrum of quantum chemistry I'm afraid, so no magic bullet. That being said, there are definitely some things I would try:

Reduce your basis set

I would second the opinion of wzkchem5 that keeping the hybrid functional is more important here than the basis set size. For the metal, triple zeta makes sense. For everything else, I would reduce to double zeta. For the prediction of electronic properties of organometallic complexes, it's still fairly common to use B3LYP with LANL2DZ on the metal (with ECPs) and 6-31+G** on everything else, which I think must go back to Yang et al. from 2009. I don't think you need to stick to that exact choice of functional/basis set, but something of equivalent size from the def2 family would make sense.

Because the metal-ligand bonding is often quite long range, it's probably a good idea to keep the diffuse functions. For the organic atoms, either ma-def2-SVP or def2-SVPD would be a good starting point. For the metal, def2-TZVP is probably fine. Make sure to benchmark your results against some higher order calculation or experimental data before committing of course.

Reduce your molecule

Hard to say for sure without knowing the structure, but you may be able to remove parts of the molecule without affecting the results. Methyls, tert-butyl groups and long alkyl chains often have little impact on the excited states, so can be removed (replaced with H). The major exception is if the alkyl group is responsible for forcing a particular molecular geometry (if sterics are important essentially), in which case the change in geometry can drastically impact the electronic states.

Try accelerated DFT

Orca has support for quite a lot of accelerating approximations that you might be able to take advantage of. Many of them are grouped under the umbrella term 'resolution of the identity' (or sometimes 'density fitting'), and they can result in big speed improvements under the right circumstances. Unfortunately (in a way), RIJCOSX is turned on automatically for hybrid functionals from Orca 5 onwards, so unless you've turned it off you're already benefiting from this speed boost. Might be worth investigating in case there's anything else you can exploit however. Have a look in the manual for RI-DFT.