Frequency calculation extremely slow

[Fransu]

I’m wondering if there is a way to speed up the vibrational analysis.
I am running a frequency calculation on a molecule containing 52 atoms for which 307 displacements are needed.
I am running the calculation on 16CPUs and 16GB memory.
The calculation run using Gaussian16 on 8CPU takes roughly 4-5h for optimization and frequency analysis but with these settings, the frequency analysis only in PSI4 1.7 using a pre-optimized structure takes more than 2 days to complete!
For both software I am running calculations at the M06-2X / Def2-SVP level of theory.

The input I’m using is the following:

memory 16 gb

molecule dummy {
0 1
*coordinates*
}

set {
    basis Def2-SVP
}

wfn = frequency('M06-2X' , return_wfn=True)
wfn.to_file('dummy_wfn')

Is there any trick to speed the process or it is a limitation in PSI4?
For a few conformers it seems very impractical to run vibrational analyses here…

Thanks for the help!

[Fransu]

PS: The molecule contains only C, H, and O atoms

[jmisiewicz]

Not realistically, no.

The heart of the problem is that we (the developers) need to implement analytic hessians for more DFT functionals. This is a project I’d love to work on, but time is not permitting.

[Fransu]

Thanks for the note! It makes sense. I have checked a bit myself and I understand that although feasible it may not be a priority.

To have analytic heassians for DFT would surely expand the applicability of PSI4 for QM studies on chemical reactions mechanisms where thermochemistry analysis is definitively required.

I hope someone will take this ‘maybe not so exciting’ task!

[jmisiewicz]

Expanding our DFT hessian and TD-DFT capabilities[1] is item #2 on my Psi development priority list. The problems are that #1 (an overhaul of our decades-old I/O system) is going to take a lot of time, and that I’m trying to finish up some research projects and a grant proposal first.

[1] = There are three “parts” of the Hartree-Fock hessian code that need to be adapted for DFT. There’s the orbital second derivative, the orbital-geometry mixed derivative, and the geometry second derivative. The orbital second derivative part is also needed by TD-DFT, which doesn’t need any other special technology. So getting TD-DFT working is not only simpler than getting DFT hessians, but it’s an excellent correctness test of one of the three major pieces. For those reasons, getting TD-M06-2X is an important stepping stone to getting M06-2X hessians. Incidentally, the path that I foresee is UKS LDA hessians → UKS GGA hessians → RKS GGA hessians → TD-UKS Meta → TD-RKS Meta → UKS Meta hessians → RKS Meta hessians

[Fransu]

Definitively! Thanks for the update and for the help - very very appreciated!

[loriab]

If you have a computer cluster and want to farm out the component jobs of the freq calc to it in an “embarrassingly parallel” fashion, Psi4 is set up to use QCArchive to do this. But it will require some extra software, setup, and management.