10
$\begingroup$

I know that CCSD(T) with the aug-cc-pVTZ basis set for geometry optimization of omeprazole (5-methoxy-2-[(4-methoxy-3,5-dimethylpyridin-2-yl)methanesulfinyl]-1⁠H-benzimidazole, the active substance in Losec) is not suitable using normal computers. The substance is:

omeprazole (omeprazole)

Can someone explain to me what exactly makes these methods too expensive for this substance?

$\endgroup$
  • 4
    $\begingroup$ Is it just that the molecule is too damn large? I've seen simple CCSD on 6-atom systems run a full week and generate 750GB of scratch work. $\endgroup$ – chipbuster Jan 13 '14 at 22:27
  • $\begingroup$ @chipbuster If you're using a computer from the early 90s, a CCSD job could take that long. However, a CCSD(T) job on 6 atoms can take just an hour with a double-zeta basis set... $\endgroup$ – LordStryker Jan 14 '14 at 1:07
  • $\begingroup$ @LordStryker wasn't my personal calculation, so I don't know. Maybe these guys were on ye olde pentium II, but more likely I'm mistaken. Thanks :) $\endgroup$ – chipbuster Jan 14 '14 at 4:15
  • 1
    $\begingroup$ @chipbuster I've run a CCSD(T)/aug-cc-pCVQZ frequency job on an 8 atom system that took about 3 months walltime (yes, core-valence basis set). Insane in the membrane. $\endgroup$ – LordStryker Jan 14 '14 at 14:53
13
$\begingroup$

To transform the question into something simpler, let's assume CISD instead CCSD(T), which is cheaper and share the common expensive step.

The molecule has 182 electrons and 1545 basis functions in the basis you specified, that means approx. 1400 virtual orbitals.

Therefore, the number of double excitation $ij \rightarrow ab$ is approximately $180 \cdot 180 \cdot 1400 \cdot 1400 = 32400 \cdot 1960000 = 63504\cdot10^6$

Unfortunately, in each of this excitation you have to calculate integral of type $(ab|cd)$, and there are $1400^4$ of them.

Such a big problem is out of reach for traditional CCSD, but tractable with more approximate methods.

$\endgroup$
12
$\begingroup$

I'd also like to add to ssavec's response. In addition to the insane amount of flops that you have to carry out to run the computation, you also run into problems of storage for the computation. In a CCSD calculation you have to store single and doubly excited amplitudes. ssavec pointed out that there are ${6.3\cdot10^{10}}$ double excitations...assuming 8 bytes per double precision floating point, that would mean storing the doubles would require 450 GB of storage! (Not to mention the I/O between iterations...)

$\endgroup$
5
$\begingroup$

Both previous answers by ssavec and jjgoings accurately describe the problem for a single point calculation. Doing a geometry optimisation, you have to run the procedure multiple times until you find a minimum on the potential energy surfaces.
Since one point already requires insane resources to compute this, an optimisation at this level will take significantly longer. From experience and common usage of chemical calculations it is often not necessary to optimise a geometry at this level. Often much simpler and more approximate methods, like density functional approximations or perturbation theory yield good and reliable results.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.