Author Topic: RI-J vs. MARI-J for 80-atom Vanadium complex  (Read 6923 times)

Tiuri

  • Newbie
  • *
  • Posts: 4
  • Karma: +0/-0
RI-J vs. MARI-J for 80-atom Vanadium complex
« on: April 27, 2007, 03:41:34 PM »
Hey,

I have a question: I am investigating a catalytic cycle of a V complex containing 64 atoms (C,H,O,N,S and V) using def2-TZVP basis set and RI-DFT BP86.
Including substrates and products, my systems can be as big as 80-100 atoms and ~2000 basis functions.

I have noticed that MARI-J is faster than standard Ri-J for my system, but now I wonder:

1.) How big is the additional error that MARI-J brings into the calculations, compared to "pure" RI-J?
2.) Can I accept the standard values for MARI-J in DEFINE or do I have to change some parameters? I read something about "f-functions" in the TM tutorial, so I think in def2-TZVP there are a lot of f-Functions, aren't there? so what about the parameters?

Sorry to sound confused :)
Thanks in advance for you help!

antti_karttunen

  • Sr. Member
  • ****
  • Posts: 227
  • Karma: +1/-0
Re: RI-J vs. MARI-J for 80-atom Vanadium complex
« Reply #1 on: May 01, 2007, 11:25:38 AM »
Hi,

1) One way to get a quick estimate of the error introduced by MARIJ in comparison to standard RI is to do two single-point calculations in the same geometry:
    1. Use only standard RI
    2. Switch on MARIJ and redo calculation. Then compare energies in the energy file.
Note that you might need to use high $scfconv in the calculations to see the effect of $marij on energy (the effect is usually very small in comparison to other approximations).

2) It depends on the version you are using. At least in 5.9 the default values for MARIJ seem already to be suitable for calculations with f-functions (lmaxmom is 14, etc.). The current revision of the manual still shows the old defaults (lmaxmom 10, etc.) that were not as strict as the new defaults.

turbomaster

  • Administrator
  • Full Member
  • *****
  • Posts: 63
  • Karma: +0/-0
Re: RI-J vs. MARI-J for 80-atom Vanadium complex
« Reply #2 on: May 10, 2007, 11:01:08 PM »
Hi,

We had to tighten the MARI-J criteria because we had ONE example (out of thousands) where
the old default values were not tight enough and led to convergence problems. It was a Pd cluster
with something like 100 atoms, which is a sick case anyway. So far, in all cases known to us
the new defaults give errors of less than 10^-8 - 10^-9 a.u., which is enough for most applications.
The computational cost increased on average by about 5% compared to the old defaults.
If you really want to push MARI-J to the limit you can try to use old defaults for a smaller system
of similar composition and see if it works ok. It should then work for larger systems as well.
We are also working on a solution which autmatically chooses best MARI-J parameters for a
given system. It still requires intensive testing and I guess it should be available in the next release.