4
$\begingroup$

It is often said that the resolution of the identity (RI) method accelerates the basis set convergence rate of MP2 "by one more zeta", e.g. that RI-MP2/cc-pVTZ/cc-pVTZ results are close to those of MP2/cc-pVQZ. If this is true, this would be very useful for my double-hybrid B2PLYP calculations (see here: Does NWChem support RI approximations in the MP2 part of double hybrids?) as both the QC packages (with Gaussian clearly distorting orbitals even for first-row diatomic molecules and Q-Chem upright refusing to show them, all calculations done on the WebMO demo server) and the NBO package (the latest version of which, used by Q-Chem, upright refusing to calculate anything with g-functions in it) seem to have serious problems with QZ sets.

My question is- is this acceleration phenomenon true in general?

$\endgroup$
1
  • $\begingroup$ +1, but when saying something like "It is often said that the resolution of the identity(RI) method accelerates the basis set convergence rate of MP2 "by one more zeta" could you provide a link to at least one reference? $\endgroup$ Commented Oct 13, 2022 at 13:14

3 Answers 3

5
$\begingroup$

Others have correctly pointed out that the "resolution of identity" method by itself is just used to help obtain approximate energies within a particular basis set. Indeed, the abstract of this paper discusses the RI-MP2 approach, and says that the RI approximation resulted in approximate MP2 energies that were within 60 milli-hartree of the true MP2 energies (over a set of more than 350 molecules). More details can also be found here.

What you're describing in your question, about stepping up by "one zeta" in convergence to the complete basis set limit, is something slightly related. For example, the abstract of this paper says that 4Z quality energies can be obtained by using a 2Z basis set when using "the R12 technology". The first paragraph of the introduction section also says that you can get 5Z quality with a 3Z basis set. The R12 technology requires evaluation of 4-electron integrals, which would ordinarily be very difficult, but are made simpler with the resolution of the identity approach (see the introduction section of this paper for example). This attempt at obtaining energies that are roughly equal in quality to the energies that you would ordinarily get with a higher-Z basis set, is also exemplified by methods such as RI-MP2-F12, which you can see is implemented as described here.

I'll point out though, that a lot of people say things like "explicitly correlated methods like R12 and F12 gain you two zetas" without understanding that this "two zeta" rule is more of a "rule of thumb" than a rigorous theorem. Observations in some datasets suggest that you can get 4Z quality energies with a 2Z basis set, or 5Z quality energies with a 3Z basis set (as described above), but this does not mean that you'll always observe a 2Z jump in quality; and the word "quality" is also open to interpretation. Some people might observe a 1Z jump (as you described in your question), and some people might observe less than a 1Z jump (i.e. barely any jump at all), and some people observe something closer to a 3Z jump in some cases. This "2Z jump" has been observed mainly for obtaining 4Z quality with 2Z basis sets and 5Z quality with 3Z basis sets, but not as much (or at all) for obtaining 7Z quality with 5Z basis sets (for example), largely because (for example) the auxiliary basis sets needed were not available to do such 5Z calculations. People have said that the higher you go up in Z, the more benefit you'll get from these explicitly correlated methods (like R12/F12), but what they don't usually seem to realize is that when you're doing calculations with a high Z (for example 5Z or 6Z), the reason is to get high-accuracy results, and the approximations associated with using these R12/F12 methods can outweigh the benefits that they provide in terms of approaching the CBS limit (and the lack of RI auxiliary basis sets for higher Z values can also be an obstacle). R12/F12 methods often in fact overshoot the CBS limit, meaning that they often give energies that are lower than the variational principle would allow. This shouldn't be a problem for you in this particular case though, since you said that you're using double-hybrids (which means that your error in the electron correlation treatment might be bigger than your basis set error anyway), but it's something to be mindful of when people tell you that you can get an 2Z jump for an arbitrary calculation by using an R12/F12 method.

I'll finally point out that since your question points out that Gaussian, Q-Chem and NBO didn't work out for you, it may be worthwhile for you to try which is available for free and very good for double-hybrid calculations, and calculations on huge systems (see here for example).

$\endgroup$
6
$\begingroup$

The RI method is an approximation, that accelerates the calculation rate but not the basis set convergence rate. So the timings of RI-MP2/cc-pVTZ may be similar to those of MP2/cc-pVDZ, but the results of RI-MP2/cc-pVTZ are nowhere close to MP2/cc-pVQZ; in fact the results of RI-MP2/cc-pVTZ are very close to MP2/cc-pVTZ, and can for almost all intents and purposes be treated as of the same quality as MP2/cc-pVTZ. Therefore, while RI can be useful for making an otherwise unaffordable MP2/cc-pVQZ calculation affordable (i.e. one can do RI-MP2/cc-pVQZ instead), it is not useful for getting rid of g functions without compromising the accuracy.

There is a similarly termed method called R12, that belongs to explicitly correlated methods, which can do what you want, i.e. attain QZ accuracy (perhaps 5Z?) using a TZ basis set. However the R12 method has largely been substituted by the F12 method, which has better cost-to-performance ratios. I'm not an expert in this field, but I think Nike Dattani knows much more than I do about explicit correlation and may have some comments on that.

$\endgroup$
2
  • $\begingroup$ Yes, for example this paper says in the abstract that "QZ quality" can be obtained using a DZ basis set, thanks to what they call "the R12 technology". The first paragraph of the introduction also says "Only a triple-zeta quality basis is needed with R12 methods to match the precision of the standard quintuple-zeta energy." I say a bit about how this is related to the "resolution of the identity" in a longer answer. $\endgroup$ Commented Oct 13, 2022 at 13:43
  • $\begingroup$ @NikeDattani Thanks! I was overly conservative about the performance of R12, because I knew that F12 typically improves the results by two zetas, and I thought that R12 is worse than F12, so maybe it only improves the results by one zeta. $\endgroup$
    – wzkchem5
    Commented Oct 14, 2022 at 15:54
4
$\begingroup$

Resolution of Identity is typically introduced to treat the 4-index integrals in your code and reduces the order of magnitude in which your calulation time scales with the number of electrons in your system, i.e. in Hartree Fock roughly from $N^4$ to $N^3$ if N is the number of particles.

It can ,as wkzchem5 already mentioned, only approximate the result you would get without RI.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .