pair style for neural network potential

Hello there,

Was just wondering if there’s any effort going into implementing a pair style for neural network potentials.

Also, if one were to do that, would it be better to do it directly as a pair style for lammps or through KIM api from a computational efficiency standpoint?

Regards,

Jay

Hello there,

Was just wondering if there's any effort going into implementing a pair
style for neural network potentials.

​you mean like this one? ​http://lammps.sandia.gov/doc/pair_agni.html

Also, if one were to do that, would it be better to do it directly as a
pair style for lammps or through KIM api from a computational efficiency
standpoint?

​the computational efficiency for whether you are using the KIM API or
program it directly into LAMMPS is for all practical purposes the same.
the benefit of KIM is that you define the pair style implementation *and*
the parameters​ *together* as a "model" and users will use it accordingly.
not to mention, that the API is available from multiple MD codes.

on the other hand, interfacing a new model is less complex when you do it
directly with LAMMPS and you have an existing template.
i would like to point out, that there are intermediate, and potentially
more effective, approaches like GAP (available via the USER-QUIP packages)
or SNAP (available via the SNAP package) in LAMMPS.

axel.