What kind of substances you consider? Which program?

Do you try ORCA?

Hi Alexey, sorry, I should have noted above. This is with Gaussian G09, and I am considering a glucose derivative in covalently bound to an amino acid sidechain. 

I have not worked with the Gaussian.

I mentioned ORCA because it has local methods allowing to work with the MP2 and even with CCSD(T) for large models.

Hello Nash,

Maybe a possibility is run a Complete Basis Set Extrapolation with MP2. It should be improved the energy from MP2 and perfermed the DFT one (more than with single-point energy). Even so, if you are sure that DFT methods describes properly the system and, furthermore your MP2 calculation is in energetically agreement with your DFT result, I think is enough. Even so, these comparisons with DFT should be done and included in the work, in my opinion.

I hope it helps you

I wouldn't expect MP2 results with such a small basis set to be any more reliable than the DFT calculation you have already performed. Something such as DF-SCS-MP2/aVTZ might be feasible with the freely available Psi4 code.

Dear Anthony,

First of all, do not benchmark any results against a MP2/small-basis-set level of theory. Smallest basis set that you should consider in MP2 calculations should be a triple-zeta, and a quadruple zeta would be perfect. We have recently compared multiple approaches for different oligonucleotides. Just for curiosity we tried out the very popular MP2/double-zeta approach and we found that the energies are sometimes very reasonable and sometimes absolutely wrong. They were over 40 kJ/mol off in some extreme cases! MP2 (as well as CC-based methods) have very slow convergence with basis set size, plus, there's additionally error coming from BSSE. So, I suggest avoiding 6-31+G(d,p) in this case.

CBS extrapolation is not possible, if you can't calculate MP2 energy with at least a triple zeta basis set. You can't do any extrapolation in this case, having not enough points...

dlpno-CCSD(T) should in principle work, but they are also risky, since it all depends on the fragmentation scheme. Often works very well, but is not as reliable as regular CCSD(T) and occassionally might be a bit untrustworthy.

The suggestion about trying out Orca is good though. It is very efficient when you employ RI or RIJCOSX schemes. You could try benchmarking your results against a double-hybrid functional like PWPB95, by Grimme. This functional has MP2 admixture, so you should also use at least a triple-zeta basis set, but I would suggest using a quadruple-zeta (we performed such calculations on even larger systems, so you should manage to run it). A single point would be sufficient, but be prepared that it might last a couple of days. For example our single-point PWPB95-D3/def2-QZVP calculations of a 120-atom-system took 3-4 days on 12 Xeon cores, per job. There are no analytical gradients for PWPB95 anyway. Remember about adding D3 dispersion correction. From your description, it seems to me that D3 will be important.

Last but not least, you use a reliable functional, so bencharking your results against another reliable DFT functional (suited for your system) is also a valuable information. 

I fully agree with Rafal. In MP2 calculations you should use at least triple zeta basis set or even quadruple zeta would be an excellent choice. Otherwise extrapolation would be far from useful. 

Regarding the codes - ORCA is a suitable choice, asides e.g. NWChem or CP2K.

Hi Rafal, thanks for your post. Point taken with regards to the smaller basis set for MP2 (thanks to Grant too). I'll be honest and admit that I'm struggling with the calculation of how much scratch space is used up during SCF calculations. At present all I know is that wb97xd/6-311++g(2df,2p) works fine (opt & freq), but when I use CCSD(T), CCSD, CC, MPn, the system runs out of scratch space (I am requesting 1500Gb!). The system currently works for CBS-4M, and I am now testing it on CBS-QB3.

I guess at this stage I need to slowly build the basis set up to see when my computational system runs out of space. 

Interesting to note, the significantly reduced basis set with MP2 returned some relative interaction energies very different to those calculated using wb97xd/6-311++g(2df,2p).

Hi Rafal, 

Just one further question. You mentioned: "

Last but not least, you use a reliable functional, so bencharking your results against another reliable DFT functional (suited for your system) is also a valuable information."

So day for example I benched mark using a different DFT functional e.g., B3LYP, would you suggest I make a comparison AFTER I redo optimisation + freq so I get the ZPE? Or, just a SPE?

Thanks

I have doubts about the 6-311++g(2df,2p). This basis gives approximately the same results as 6-311+g(2d,p), but it is significantly more expensive. If 6-311+g(2d,p) is not accurate enough, I would go directly to aug-cc-pVTZ but not to 6-311++g(2df,2p).

Hi Alexey, 

I have no experimental data in this particular compound to compare with so saying whether something is accurate enough it proving difficult. I have suggestions from papers based on very similar compound which use wb97xd/6-31+g(d). Unfortunately I can't use counterpoise to account for BSSE, so I used as big a basis set as I could (perhaps not big enough on account of you message).  It is a compound in human tendon that is a) theoretical in structure, and b) proving extremely difficult to synthesis. It's presence has been identified, just not the structure. 

6-311+g(2d,p) is a significant improvement over 6-31+g(d). And it provides relatively small superposition error among the basis sets of the same size. It's very subjective, but I consider this as the last useful Pople basis.

Thanks for the insight Alexey. I've decided to stick with wb97xd/6-311++g(2d,p) for optimisation + freq + energy, but I am comparing it to SPE of mp2/augccpvtz and mp2/6-311++g(2d,p), in addition to SPE of CBS-QB3.

What is the comparative computational cost between MP2/RI-MP2 and SCS-MP2?

Anthony Nash Alexey Nikitin Rafał Szabla

It's exactly the same.