atom types for coarse graining

[KP-303]

If I want to use allegro to learn a coarse grained potential, how do I assign atom types in my dataset? Nequip insists on knowing the type of every atom. I could just lie and call them all carbon, but will this interfere with the training? Just checking there isn't something in the background that depends on atom type.

Thanks

[DavidW99]

No, it should not. The atom type is encoded by one-hot encoding. It gives a vector with the length of the total number of atom types and it is 1 at the specie's index and 0 elsewhere. You will assign the chemical symbol to the type index in the yaml file right after you specify your dataset.

# A mapping of chemical species to type indexes is necessary if the dataset is provided with atomic numbers instead of type indexes.
chemical_symbol_to_type:
  C: 0

Therefore, NequIP will not know your atomic type but only the type index that you assign.

[Linux-cpp-lisp]

Note also that you can just provide arbitrary type names, in order, instead of chemical_symbol_to_type:

type_names:
 - Type1
 - SomethingElse

where here type 0 will be named Type1, etc.

In this case, however, the input data must include an atom_type field instead of atomic_numbers. In the extxyz format, this can be done with a custom per-atom atom_type field, and just giving X as the "species" (atomic number 0) for the atomic numbers of all atoms, for example.

I think the ZBL potential should be the only provided module that depends on the atomic number right now in a physical way.

[KP-303]

OK thanks.My beads can have a charge and still be the same bead type as the uncharged version. I have included a charge column in my extxyz training data. How do I make Nequip use the charge as a descriptor in the training? Or should I just relabel charged beads to be a different 'atom type'?

[Linux-cpp-lisp]

If you have a set of discrete particle types that just happen to be distinguished also by some notion of charge here, yes, relabeling them as different "atom types" would be reasonable.