Using Machine Learning model for only intramolecular interactions

Hello LAMMPS representatives and users,

I have a question, I hope this is the right place for it:
Say I want to simulate the system consist of graphene or hBN and some molecule (say benzene ring based small molecules) on top of it. I want to use trained machine learning model just for molecule and layer instead using intramolecular force fields for these molecules which can be represented by Tersoff or REBO for example. For intermolecular interactions, I am going to use Lennard-Jones (I know it is very tricky because of the huge dislocated electron could on the surface of graphene) or say ILP potentials which I recently found. My question is that is it technically possible?
And one more important question is of course that can this approach give a realistic results?
Because I prepare ML model for molecule in isolated case, or should I prepare it using the same system that ML model trained in the case of intermolecular interactions included (at the end of day I use other interlayer potential)?

I appreciate any suggestions.

Best Regards,
James.

Yes, this is “mechanically” possible, if you use pair style hybrid and all pair styles are compatible with it (not all machine learning potentials “in the wild” are).

That will strongly depend on the specifics of the system and what kind of results and level of accuracy you are expecting. It is quite possible that the approximation from using a hybrid/mixed potential approach can cause a larger error than creating a machine learning potential for the entire system. But then again, there is no guarantee that the situation is not the inverse.

My advice would be to try and construct a set of sufficiently small tests and use them with different combination of models to see what will give consistent results and try to relate those to available experimental data. By using a hybrid force field, you are effectively creating a new force field and thus will need to validate it just like any other new force field.

1 Like