Support of external NN potentials

[mmagithub]

Hello,

I am wondering if it is possible to use NN potentials for energy evaluations to get faster results for computationally exhaustive computations, like torsional scanning, something like ANI-1x or AIMNET:

https://github.com/aiqm/aimnet

Or this is already possible now ?

Thanks,
Marawan

[selimsami]

Hi Marawan,
Currently not implemented but it would not be quite straightforward to add a new QM software to use NN potentials if you'd like to contribute.
Alternatively we're very close to finishing an update where you can specify different QM methods for different steps, so that should bring the computational cost down.
For example you can do the torsional scans with xTB and do a single point calculation at the end with a higher level of theory.

[mmagithub]

Thanks Selim, this will be a very good addition to this magnificent project.

[xiki-tempula]

Hi @selimsami ,

Alternatively we're very close to finishing an update where you can specify different QM methods for different steps, so that should bring the computational cost down.

I wonder how close we are at finishing this? I'm found this part of the calculation cumbersome as well. One could also do a constrained GeoOpt after the xTB run to get the geometry optimised at the desired QM level at a much cheaper cost.

[selimsami]

Hi Zhiyi,

We just finished coding it in the xtb branch, still in the testing process though. Feel free to give it a try and provide feedback or wait another week till it's merged.

We did add a "preopt" option as well so that you can indeed do a pre-optimization at a low cost. Not sure what you mean by constrained though - what do you want to constrain?

[xiki-tempula]

So what I commonly do is that you have the QM GeoOpt conformer from your hessian calculation.
Then I run xTB torsiondrive to get the xTB GeoOpt conformer at each torsion point (24 conformers if use 15 degree spacing). One can then run single point calculation on these conformers. However, the downside is that the conformer is not exactly the GeoOpt conformer at the specified QM theory.
Thus, what I usually do is that I use these 24 conformers and input to the QM engine and go GeoOpt while constraining the torsion such that all other degree of freedom are relaxed and thus, one get the GeoOpt QM conformer at the selected level of theory.

[xiki-tempula]

I had a brief look, as it is a very large PR. I think in principle, I'm happy with the changes.
I could fill in the ORCA bits when this PR is merged.

[selimsami]

Hmm I see. We did not add the option to do a pre-optimization for the dihedral scans, but what you can do now with all software is do a single point QM with a software of your choice after the torsiondrive. It wouldn't be too complicated to replace that with a constrained optimization I guess.