Can I freeze f functions from a TZ basis set during MP2 (but not during H-F) to speed up calculations?

Question

This is another question about my molecule; as I'm not doing this as an academical affiliate(I'm currently majoring in mathematics), I have only NWChem and my personal laptop on the mix, and thus need to save time and/or CPU usage.

I noticed, already on the WebMO demo server, that DFT calculations on elements no larger than Ar often fail at the 45-second CPU time limit (de jure 30, de facto 45) on said server if they contain f functions in their basis sets.

WebMO Basic and NWChem are both free without any official CPU time limits- the problem is that I use my laptop to code things (that are completely irrelevant to computational chemistry) already and thus do not want to "torture" my laptop unnecessarily more.

I mentioned f (polarisation) functions on both the title and the body of the question- the desired question and its reasons are now clear.

UPDATE: My question was originally about freezing f-functions from the TZ basis set ONLY in the MP2 part of the calculations; i.e. doing a full TZ H-F calculation AND THEN freezing the virtual f-functions in the forecoming MP2 calculation(See MP2 page in the NWChem manual to see what I'm talking about). The question somehow got changed to something entirely different so I'm adding this here for more clarity.

Answer 1

It depends on what you're intending to do:

• If you are comparing to benchmark results with this basis set, then likely no.
• If accuracy is very important to you, then likely no.
• If you are intending to do a basis set extrapolation, then it might compromise the quality of your extrapolation, although Jacek Koput took the k-type functions out of the aug-cc-pCV7Z basis set before doing extrapolations, and those extrapolations still gave excellent results.
• If you don't care too much about accuracy, then sure, remove the f-type functions.

It all depends on the level of accuracy you want, and what price you're willing to pay to achieve that level of accuracy. You can remove some f-type functions, and you'll most likely get slightly less accurate results because your basis set is smaller and offers less variety, but your results will still likely be far more accurate than what you would get by reducing the basis set by an entire zeta value.

If you have enough computer power and you care about accuracy, it would be best to include the f-type functions. If you don't have enough computer power, there's nothing wrong with removing the f-type functions as long as you do it for all energies involved in the energy-difference calculations, so that you get "parallel errors".

It may also be a good idea to do some tests to find out how much of an impact those f-type functions are making. Perhaps find a similar molecule that's a bit smaller, so that you can do calculations with and without the f-type functions, and see how big of an impact this has on the final energies. The difference in those energies can be incorporated into your analysis of the uncertainty for your final calculations in the system in which you're truly interested.

Answer 2

The answer is no. Post-Hartree-Fock methods are famously slow to converge to the complete basis set limit due to their dependence on the unoccupied orbitals. Modern basis sets, such as the Dunning cc-pVXZ series and the Karlsruhe def2 series are designed to achieve the best compromise of accuracy and performance. By freezing the $f$ orbitals, you compromise this accuracy, especially since the $f$ orbitals are more important at the MP2 level than the HF level!

I would also like to point out that while NWChem is not known to be competitive on single-node performance; its focus is on massively parallel calculations. As we have recently discussed in our open access review in Wiley Interdiscip. Rev. Comput. Mol. Sci. 12, e1610 (2022), there is a wide variety of free and open source codes that can be easily installed on personal hardware and be used for realistic computational chemistry.

Psi4 and PySCF are two personal favorites of mine: Psi4 is easy to use while PySCF offers a powerful toolbox for designing new methods. Psi4 is especially attractive for its focus on choosing fast algorithms by default: it makes using density fitting methods easy, which can significantly speed up MP2 calculations, for instance. Psi4 also implements frozen natural orbital methods, which are a safe way to economize coupled-cluster calculations by freezing out natural orbitals that have small occupations at the MP2 level of theory. Psi4 also has built-in support for basis set extrapolation, which likewise offer an easy way to economize on the cost of calculations.

Answer 3

I'm not sure if freezing at the basis set level makes sense... However, you can make an MP2 calculation easier, by doing an active space approximation (freezing certain orbitals). You will need to have manual control going from the atomic orbitals to molecular orbitals.

All you need to do is after the Hartree-Fock calculation, when you convert the atomic orbitals to molecular orbitals, you then need to make an active space approximation (find the core constant energy, arising from the occupied frozen orbs then remove them).

I don't use NWChem, but PySCF will do let you do this. See (subsection Frozen orbitals):

https://pyscf.org/user/mp.html

I'd just use this, as you won't have to reinvent the wheel then :)