Add PolarMACE electrostatic ML/MM embedding via graph-longrange

[CheukHinHoJerry]

Draft — opened for early feedback on the approach after asking claude to refactor the implementation, not as a finished PR. Trying to adapt to some convention from EMLE too. Two questions:

• Suggestions on tests that I should be adding?
• There are a huge portion of the code here spent on checking validity of the MM forcefield but I saw EMLE implementation not doing all these yet. How should I go with this?

Summary

Adds PolarMACE electrostatic ML/MM embedding to the MACE potential, using the reusable external-source APIs from graph-longrange. The MM field enters the model's descriptor and polarises the ML density inside a single forward pass, and the MM back-reaction forces are returned and scattered into the full force array.

Dependency

Requires the external_field branch of graph-longrange, which provides the external-source energy and feature blocks:

pip install 'git+https://github.com/WillBaldwin0/graph_electrostatics.git@external_field'

devtools/requirements/mace-aimnet-torchmd.txt is pinned to that branch which supports electrostatic embedding by allowing PolarMACE taking external point charges input.

How it fits the existing structure

MACEPotentialImpl.getSupportedEmbeddings() advertises "electrostatic", so MLPotential.createMixedSystem() routes that name here and leaves "mechanical" to the generic embedding plugin. addForces() still handles both, since the generic plugin calls it.

Minimal example on how to use the code

from openmm import unit
from openmm.app import PDBFile, ForceField, Simulation, PME, HBonds
from openmm import LangevinMiddleIntegrator
from openmmml import MLPotential

# load system
pdb = PDBFile("system.pdb")
# MM forcefield setup
ff  = ForceField("amber14-all.xml", "amber14/tip3pfb.xml")
mm  = ff.createSystem(pdb.topology, nonbondedMethod=PME, constraints=HBonds)

# The ML subset — the QM region.
mlAtoms = list(range(120))

# using mace potential, can also be some trained MACE-POLAR models
potential = MLPotential("mace-polar-1-small")

# Only PolarMACE-class models advertise it; plain MACE returns [].
assert "electrostatic" in potential.getSupportedEmbeddings()

mixed = potential.createMixedSystem(
    pdb.topology, mm, mlAtoms,
    embedding="electrostatic",
    charge=0,            # ML-region total charge
    multiplicity=1,
)

# standard NVT dynamics
sim = Simulation(pdb.topology, mixed,
                 LangevinMiddleIntegrator(300*unit.kelvin, 1/unit.picosecond,
                                          1*unit.femtosecond))
sim.context.setPositions(pdb.positions)
sim.step(1000)