Future GFN2-xTB implementation?

1

Hi all,
I know Psi4 does not currently support any semi-empirical methods but are there any future plans to implement the extended semi-empirical tight-binding (GFN2-xTB) model in Psi4?

Thank you,
David

2

Hi, I think there is interest but its on no ones schedule afaik.
It can be relatively quickly implemented via the QCEngine interface.

1 Like
3

I am excited to hear that there is interest and that the implementation would be relatively painless! Specifically, we are interested in running a GFN2-xTB calculation and obtaining the output wavefunction as a Psi4 object for further processing. We would like to incorporate GFN2-xTB outputs with an ML model that we have developed built around Psi4 formatted wavefunctions. Would running GFN2-xTB with the QCEngine accomplish this? More generally, we are interested in any method that could convert a GFN2-xTB wavefunction to a Psi4 object to integrate in our Psi4 workflow.

4

It would essentially be an external program for psi4. I am not sure about the extend of the QCEngine harness for xtb in terms of returning e.g. Fock matrices, densities etc
I fear that is not available out of the box. Even through the python API from xtb itself (API Documentation — xtb-python documentation)

I see only energies, gradients, dipole and point charges in the test suite in the QCEngine test suite.
What data from the wfn is needed?

5

A psi4 native reimplementation is fairly unlikely, but the developers of xtb are very active and if some additional functionality is needed one could make it work together.