In order to use a calculator for a Force field that is not included among the standard ones (e.g. the original chgnet) the suggested solution is to pass

    {
        "@module": "chgnet.model.dynamics",
        "@callable": "CHGNetCalculator"
    }

to the force_field_name in the subclasses of ForceFieldMixin.

The problem is that, while this works with the generated Job object, it fails when the object is serialized. The following example:

from atomate2.forcefields.jobs import ForceFieldRelaxMaker
from jobflow import run_locally
from pymatgen.core import Structure
from monty.json import MontyDecoder, MontyEncoder
import json

mgo_structure = Structure(
    lattice=[[0, 2.13, 2.13], [2.13, 0, 2.13], [2.13, 2.13, 0]],
    species=["Mg", "O"],
    coords=[[0, 0, 0], [0.5, 0.5, 0.5]],
)

flow = ForceFieldRelaxMaker(force_field_name = {"@module": "chgnet.model.dynamics", "@callable": "CHGNetCalculator"}).make(mgo_structure)
flow = json.loads(json.dumps(flow, cls=MontyEncoder), cls=MontyDecoder)
run_locally(flow)

fails with

ValueError: Could not create ASE calculator for MLFF.Forcefield.

This leads to a failure in jobflow-remote, that needs to serialize the job in the DB, but I suppose the same would happen with Fireworks.

The primary reason is that the calculator_meta is set at creation time, but not serialized due to the presence of field(init=False)

Simply removing the init=False does not solve the issue though, because upon deserialization of the Job the {"@module": "chgnet.model.dynamics", "@callable": "CHGNetCalculator"} is automatically deserialized by the MontyDecoder and the code only expects either an MLFF or a dict for calculator_meta.

I managed to make it work by allowing calculator_meta to also be a Calculator subclass. I can open a PR if this seems a reasonable solution, but maybe there are better options?