We'd like to present our dipole assessment with different charge models in the next openff release call and I found something really really strange with espaloma. I'd like to give priority to this at the moment.
LD – BSwope and DMobley and I are looking at dipole quality using different charge models. Comparing AmberTools AM1BCC, OE AM1BCC, OPLS+-vsites, Espaloma, and Nagl. The results for the first three are pretty consistent, but the dipoles for the latter two are really different. We’re looking at 3 metrics comparing MM to QM dipole (b3lyp)
Angle of dipole vector
Length of dipole vector
magnitude of dipole vector
The GNN results looks really off, I think because there’s an issue with atom ordering becoming misaligned with charge ordering.