Compare timings for ASE/xtb-python versus xtb

[danielhollas]

If we want to preoptimize molecules by xtb in real time in SMILES widget, it needs to be fast. And currently, it is not fast even for smallest molecule.

It seems that the problem is the slow ASE interface, but we need to verify that.

See issue

https://github.com/grimme-lab/xtb-python/issues/38

Unfortunately, optimizing directly via xtb might be non-trivial, per:

https://github.com/grimme-lab/xtb/issues/403

Perhaps as an ugly workaround we could ditch xtb-python and shell directly to xtb and parse the results?

Note also some additional info in https://github.com/aiidalab/aiidalab-widgets-base/issues/161

[awvwgk]

And currently, it is not fast even for smallest molecule.

Fast in the sense of performance for the single point calculations or in the sense of iterations needed in the optimizer? The latter is were ASE is lacking compared to the optimizer in xtb.

[danielhollas]

@awvwgk good question. My impression is the former. Even just two iterations using ASE optimizer takes several seconds for small organic molecules (couple carbons). Perhaps this is expected, perhaps I am doing something wrong. Either way I need to do a proper benchmark, hopefully sometimes in the next two weeks.

I can test the overhead of xtb-python/ASE by directly calling xtb from Python.

[danielhollas]

https://github.com/jvalegre/aqme/commit/2799ee815a3ac91623c15dafc8a6f3eb8a7e649b

[danielhollas]

We might be able to use tblite with ASE, in case that is faster https://tblite.readthedocs.io/en/latest/users/ase.html