How important is it that geometry be optimized at a high level of theory?

Question

I've been doing research with a computational chemist for a little while now and in one of our projects we are dealing with a rather large system and because I go to a relatively small university, our resources are limited. Thus, when we began to ramp up the calculations in the project, we opted not to re-optimize our structure at CCSD(T)/def2-ATZVPP despite the fact we are doing the rest of our calculations at that level of theory with that basis set. The reason we chose not to optimize is we have a limited amount of time/resources and we estimated the optimization would take around 45-50 days to complete which would basically hog a computer for that long and stall the other projects we are working on.

I'm wondering how significant of an error, if much at all, that will introduce?

To give a few more details, we are studying vibrational properties of the system using a local mode approach. The structure we have is optimized at MP2/def2-ATZVPP.

Regardless, I'm just looking for a general understanding of how important the optimized structure might be. It seems like it would be less important, but could be quite dependent on what one is studying, than other things based on the fact that no system is truly at that optimized structure, but is only in that geometry on average. Is that correct at all?

Answer 1

Method

• Most of the time CCSD(T) would indeed be a huge overkill for geometry optimisation. DFT and MP2 have way better performance/cost ratio. Note that I said DFT and MP2 above, not or MP2; this is well known procedure to compare DFT geometries with MP2 ones to be on the safe side: there should be no big differences.
• On the DFT side hybrids might be overkill as well. Try simple pure GGA functionals like BP86 and PBE as well meta-GGA functionals like TPSS first. There might be almost no advantages in using hybrids for geometries.
• Include empirical dispersion corrections in DFT calculations even if you do not suspect dispersion forces to be decisive. Just in case.
• Employ the density fitting (also known as the resolution of identity) for DFT as well as MP2 to significantly accelerate calculations.

Basis set

• With respect to the basis set convergence of geometries is typically reached at the triple-zeta level. This is fairly well-known from excessive benchmark studies for correlation consistent¹ and polarisation consistent basis sets², but Turbomole def2 basis sets should have similar behaviour as well. References are exclusively about DFT, but the same is true for WFT as well.
• In practice, though, for DFT calculations (but not necessarily for MP2) a triple-zeta basis might also be an overkill. Try double-zeta def2-SVP first (augmented with diffuse and polarisation functions to your taste). Even Turbomole developers themselves describe DFT/SV(P) and MP2/TZVPP levels of theory as "almost quantitative".

Benchmark

• Compare the calculated geometries with experimental data if such are available for your (or at least similar) compounds. At the end of the day, how well does a particular level of theory reproduces experimental geometries is the main criteria of choice.

---

1) Wang, N. X., & Wilson, A. K. (2004). The behavior of density functionals with respect to basis set. I. The correlation consistent basis sets. The Journal of Chemical Physics, 121(16), 7632–46. http://doi.org/10.1063/1.1792071

2) Wang, N. X., & Wilson, A. K. (2005). Behaviour of density functionals with respect to basis set: II. Polarization consistent basis sets. Molecular Physics, 103(2-3), 345–358. http://doi.org/10.1080/00268970512331317264