Can not make openmm system with no `Nonbonded` force

Describe the bug Trying to create an openmm system with no nonbonded force results in an error due to the post-processing done here. In my case the nonbonded force is replaced by a custom nonbonded force implemented via smirnoff-plugins which creates the exclusions here before the toolkit so missing this step should be fine.

To Reproduce This is not quite the same as I have not added the custom force but the result is the same

from openff.toolkit.typing.engines.smirnoff import ForceField
from openff.toolkit.topology import Molecule

ff = ForceField("openff_unconstrained-1.3.0.offxml")
# remove all nonbonded handlers 
del ff._parameter_handlers["ToolkitAM1BCC"]
del ff._parameter_handlers["LibraryCharges"]
del ff._parameter_handlers["Electrostatics"]
del ff._parameter_handlers["vdW"]

water = Molecule.from_smiles("O")
system = ff.create_openmm_system(water.to_topology())

Output

IndexError Traceback (most recent call last) /var/folders/9q/nm__l0v13fggc72v94p0hybw0000gq/T/ipykernel_68062/4179821129.py in <module> ----> 1 system = ff.create_openmm_system(water.to_topology())

~/miniconda3/envs/openff-plugins/lib/python3.8/site-packages/openff/toolkit/typing/engines/smirnoff/forcefield.py in create_openmm_system(self, topology, **kwargs) 1376 # TODO: Can we generalize this to allow for CustomNonbondedForce implementations too? 1377 forces = [system.getForce(i) for i in range(system.getNumForces())] -> 1378 nonbonded_force = [f for f in forces if type(f) == openmm.NonbondedForce][0] 1379 1380 nonbonded_force.createExceptionsFromBonds(

IndexError: list index out of range

Computing environment (please complete the following information):

• Operating system osx
• Output of running conda list
• openff-toolkit 0.10.0+32.gc316bfe0.dirty
• openff-toolkit-base 0.10.0
• smirnoff_plugins 0.0.1+4.g6118a88.dirty

Additional context

[mattwthompson]

The TODO two lines above seems quite relevant as a bit of debt 😅 .

Relates somewhat to discussions around revamping the SMIRNOFF spec, particularly https://github.com/openforcefield/standards/issues/2#issuecomment-930133340. Requiring at least one of NonbondedForce or CustomNonbondedForce seems to fit every use case I can think of. Any obvious pitfalls?

[SimonBoothroyd]

I agree that, though peculiar, force fields with no non-bonded interactions should be permissible.

When computing gradients by finite difference I usually create a subset of a force field that only contains the relevant handler type. So if I was computing dU / d bond length then I would create a system from a force field only containing the BondHandler. Kind of a weird use case but definitely useful to be able to do.