RI for hybrid DFT

[Marisa]

Hello everyone,

  If I want to run a RI approximated hybrid DFT calculation,

(i)Is their only serial calculation available now?
(ii)There are RI-J, RI-KJ, shall I use both of them in order to get maximum efficiency?

Thank you~

[uwe]

Hi,

i) yes, RI-hybrid DFT is currently only available in the serial version.

ii) just RI-J, so do NOT switch on rijk in the last menu of define, just ri.

RI-JK is faster than non-RI calculation only for mid-sized molecules with large basis sets. It has mainly been developed as a pre-step for RI-MP2 and RI-CC2 where large basis sets are the default.

If RI-JK is used for small basis sets like SVP, it can be much slower than the non-RI or the RI-J calculation!!

Regards,

Uwe

[Marisa]

Thanks for your explanation, from the manual
  http://www.turbomole-gmbh.com/manuals/version_5_10/DOK_HTML/node266.html
RI-JK can reduce the cost of HF, therefore may reduce the HF part of hybrid DFT?

From your post, seems actually RI-JK is not so helpful in many cases, did I understand correctly?

[uwe]

Hi,

RI-JK is very helpful in cases where you have larger basis sets - this is also described in the documentation that you have cited yourself: [ ... ] (efficient for large basis sets) [...].

If you are using at least a triple-zeta basis function plus polarization like TZVP, TZVPP, QZVP, etc. RI-MP2 calculations will be much faster than the Hartree-Fock run you have to perform first. RI-JK in that case is very efficient and reduces the time by an order of magnitude.

My rule of thumbs are:

• Hartree-Fock, MP2 or CC2
  
  • use basis sets of at least TZVP quality (6-311G* in Pople basis set nomenclature)
  • use RI-JK - if the number of basis functions is not too high (depends on the amount of memory you have, roughly 1000-1500 basis sets is usually the limit where the non-RI calculation gets more efficient again).
  • the more memory you have and can assign with $maxcor the better the performance will be (more or less)
• DFT
  
  • use basis sets of SVP or TZVP quality ( 6-31G* or 6-311G* in Pople basis set nomenclature)
  • use RI-J
  • use MA-RI-J (switch on marij in the last menu of define) - this is always safe and speeds up the calculation of larger systems by another factor of 2-10.

It is of course not forbidden to use any other combination like MP2 with small basis sets or RI-J with large basis sets, etc. But if you want to perform calculations for real-life applications in a reasonable time with reasonable results, the combinations described above should be applicable in most cases.

Regards,

Uwe

[Marisa]

Hello,

  Many thanks for your reply. After some test, for my case, B3LYP/def2-TZVPP with 1500 basis function, RI-JK and RI-J together are even slower than RI-J alone

Best regards