Moderators: Global Moderator, Moderator
Hello,
Is there a routine for combining MLFF with electric field calculations? I have tried applying electric field during on-the-fly training, and the polar molecule aligns with the applied field as expected. However, during production runs the molecule does not align with the applied field. I have checked the forces in ML_AB with training without electric field and they are different, but during the production run the behaviour is the same.
Regards
Sindre
Hello Sindre,
I am sorry but as far as I know there is currently no way of combing machine-learned force fields with electric fields in VASP. As you already mentioned the reference forces ending up in the ML_AB database will be different, so there is some influence on the training. However, there is a conceptual problem which prevents that the trained differences are correctly applied: in a first step the ML force field method transforms Cartesian coordinates into translationally and rotationally invariant local environment descriptors. Hence, if you would apply a global rotation or translation to the structure you would get exactly the same energy and force predictions. Usually, this is a big advantage because there is no need to learn invariances from the data. On the other hand, it poses a problem if we actively want to break these symmetries by applying an electric field along a certain direction. Without modification there is simply no "input" to the ML force field method which would transfer the direction and strength of the electric field so the force prediction could take it correctly into account.
All the best,
Andreas Singraber