Long range training: Benzene ring problem

[Florian-Pabst]

Dear all,

I am trying to to train a Deep potential long-range (DPLR) model on a molecule containing a benzene ring.
Since the electrons in the ring are delocalized, I can not unequivocally assign Wannier centers to atoms. Therefore, I calculated a Wannier centroid for the whole molecule, like it is done in the water example. However, since in my case both atom types (C and H) are present more than once per molecule, this centroid can not be assigned to a single atom, like in water. So I trained the dipole model to learn the distance of the centroid from each C atom. This results in multiple parts (as many as C atoms) of this centroid with non-integer charge, very close to each other, in the DPLR training. I guess this is where the problem lies, because I obtain huge training errors, but I can not think of an other way of doing it.

Thanks in advance for any suggestions!