How to use correctly the pair style "hybrid/scaled" with lookup tabel?

LAMMPS LAMMPS General Discussion
1

Dear LAMMPS community,

I have the following scenario:

  • I have a 2D sheet of graphene, that I cut into a disk shape-like, then I terminate the edges of the sheet by placing hydrogen atoms at a distance of 1.08 Å.

  • To run my simulations, I want to use two PACE-MLIPs: the first is carbon-based and the second can treat hydrocarbons.

    • MLIP specifications:
      • PCAE-carbon based (Qamar) : trained separately on DFT-pbe with the advantage to turn off D2 dispersion correction if not needed (additional loockup table of D2 correction is provided that can be used by invoking hybrid/overlay and the sub-style table, via pair_coeff * * ./d2.table )
      • PCAE-hydrocarbon based (Willman) : Trained on DFT-pbe+dispersion correction (no separate loockup table to turn on/off the dispersion).

  • Being inspired by the feature of Type Label Framework for Bonded Force Fields in LAMMPS, I subdivided my atoms into 3 types:

    • C-center: The carbon atoms near to the center of the mass within a defined radii,
    • C-ext: The carbon atoms that come just after the core region atoms,
    • H-edge: Hydrogen atoms that decorate the my structures.

So, my goal is be able to run them both while:

  • Core atoms are assigned to PCAE-carbon based + D2 correction;
  • Extended carbon atoms use D2 correction, in additions to a weight of 1/2 for both PACE-MLIP (via hybdrid/scaled);
  • Hydrogen atoms are maintained by PACE-hydrocarbon;
  • The interactions between extended carbon atoms and hydrogen atoms are hold using PACE-hydrocarbon;
  • Finally, one can ignore carbon core atoms and hydrogen atoms, or leave it to the hydrocarbon PACE.

I am confusion about how to achieve this or if is it even doable? because the pair_style is defined once and for the entire simulation, and what I understand is to be able to use table sub-style we can use hybrid/overlayed, and If I want to have a weighted action on certain atoms I need to use hybrid/scaled.

LAMMPS version: LAMMPS (22 Jul 2025 - Update 1)

Thank you in advance.

Best regards,

Anouar

2

As has been discussed here many times before, piecing together a model from multiple many-body potentials is rarely a good idea and your suggestion is among the worst of those.

In a hybrid model, each sub-system is computed separately without any knowledge of the presence of the other atoms and thus all cross-terms must be provided in addition and those can only be realized by pair-wise additive potentials. Thus any gains from using machine learning potentials will be lost by having to use something like lj/cut or morse for those. Furthermore, using two different parameterizations for atoms in the same compound would be inconsistent and thus a bad idea, even if the cross-term requirement would not exist.

Your suggestion for adding the dispersion correction term is even worse, so I don’t even want to discuss it.

TLDR: you should use either one single machine learning potential parameter set for the entire system or use a single traditional classical force field like OPLS-AA. The latter will likely be a better representation of the system - despite all its limitations - than the “chimera-zombie model” that you are proposing. It is thus pointless to even try to figure out the “mechanics” of making LAMMPS run.