Differentiable classical force field support

[WangXinyan940]

We propose to add the following features to DMFF:

• 'LennardJonesForce'
• 'CoulombPMEForce'
• 'CoulNoCutoffForce'
• 'CoulReactionFieldForce'
• 'HarmonicBondJaxForce'
• 'HarmonicAngleJaxForce'
• 'PeriodicTorsionJaxForce'

The following generators are added to support related cards in OpenMM type XML forcefield.

• 'HarmonicBondJaxGenerator'
• 'HarmonicAngleJaxGenerator'
• 'PeriodicTorsionJaxGenerator'
• 'NonbondJaxGenerator'

[Ericwang6]

Here should be solved

[KuangYu]

See comments in code.

[KuangYu]

We have the same constant in dmff.admp.pme, called DIELECTRIC, should we consider put all important constants together? Maybe add a constant.py file under dmff.

[KuangYu]

Seems like in the classical module, all positions/box are assumed to be in unit of nanometer. But in ADMP, all positions are assumed to be in Angstrom. We should discuss this unit conflicts...

[WangXinyan940]

I'd like to input position/box in nm & kJ/mol in API, and then transfer to angstrom when creating potential function like this:

class ADMPDispGenerator:
    def createForce():
        def potential_fn(positions, box, pairs, params):
            position_ang = positions * 10.0 
            box_ang = box * 10.0
            E_sr = pot_fn_sr(
                position_ang, box_ang, pairs, mScales, a_list, b_list, q_list, c_list[0]
            )
            E_lr = pot_fn_lr(position_ang, box_ang, pairs, c_list.T, mScales)
            return E_sr - E_lr

Then we don't need to touch parameter preparing process.

[KuangYu]

This is good.

[KuangYu]

This may cause precision issues. The way in admp/pairwise.py might be better? Try to apply scaling factor directly on pairwise interactions.

[WangXinyan940]

OK. Making change now.