0
Performance
Viewed 33 times

Hi Jan,

You state on page 12 that MOPAC was faster than GAMESS to converge (despite taking more steps) - was this due to the use of the MOZYME localized molecular orbital (LMO) method?

If not, what effect on the speed and accuracy of MOPAC in your comparison do you think it might have had?

You also state on page 14 the modest increase in speed seen when running the programs across parallel processors. Do you think the "fragmenting" approach of MOZYME ought to assist with this problem?

I thought it interesting that your finding that including three-body dispersion correction had no substantial effect on accuracy as this is also what Li & Muddana found last year. Can you make any guesstimate as to where your PM6-DH3+ method might have placed on their Time vs RMSE comparison graph (they only compare the older PM6-DH+ method)?

Many thanks, Doug

waiting for moderation
1 Answer
0
Accepted answer

Hi Doug

We didn't use Mozyme/LMO. There is just a bit more overhead in GAMESS, which becomes apparent when you start doing semi-empirical methods. The overhead is insignificant for methods that are more costly in comparizon, like Hartree-Fock, DFT. I think the MOPAC diagonalizer is also substantially faster, although I am not sure how they differ exactly.

Better parallel scaling would require implementation of a better parallel diagonalizer, I am unsure if Mozyme/LMO's would benefite here. Definitely something like the Fragment Molecular Orbital (FMO) method would speed things up in this regard.

There are not any molecules in the training data set that have substantial three-body dispersion effects. This would likely require a different set of systems to elucidate any significant interactions. Maybe something like the L7 data set would be suffficient, which has three-body interactions on the order of 1-2 kcal/mol.

I am not sure which Li&Muddana paper you are refereng to. The speed of thePM6-D3H+ and PM6-DH+ methods are virtually identical (if implemented the same way), and I guess the accuracy in terms of electronic energy boils down to how well the -D2 or -D3 corrections are able to describe the dispersion interaction, which certainly varies a lot from system to system.

/Anders

waiting for moderation