Neural Network Energy Potential with Tabulated interaction functions

GROMACS version: 2020.3

Hi everyone,
I would like to use a force field in GROMACS coming from my Trained Neural Network. My Neural Network function represents the local potential energy of an atom i , E_{i}.

E_{i} depends on, say, n descriptors (D_{i1},…,D_{in}), E_{i}(D_{i1},…,D_{in}) . These descriptors depend on coordinates of atom i and coordinates of the first, say, 20 nearest neighbour atoms.

I found the Tabulated interaction functions topics in the reference manual. It seems possible to use a custom pairwise interaction E(r_{ij}) but not more.

I’m looking for a routine where:

I may compute at each step the descriptors for each molecule than evaluate all E_{i} and then derivate it to get forces.


I may tabulate all the descriptors and the corresponding energies and forces.

Is there a way to do something like that in GROMACS?


No, this is not possible in GROMACS.

We used to have support for implicit solvent, which had pair potentials that used coefficients based on local densities, which depended only on (local) distances. It sound like what you need is more complex than that, otherwise you could have a local at the old implicit solvent code.

Thank you for your suggestion. I’will look at the implicit solvent code.