[DNM] add espaloma converter ...

[lilyminium]

Is this suitable for mapping the espaloma experiments to smee/descent/dimsim ...?

Just prototyping code, wouldn't suggest merging this while espaloma is under heavy development.

This basically works by assigning each bond/angle/... its own unique types. It assumes that torsions have 4 cosines but you can set that.

[lilyminium]

For python formatting/highlighting, could look like below:

dhmix_property = DataEntry(type="hmix", smiles_a="CCO", smiles_b="O", x_a=0.5, x_b=0.5, ...) # mapped smiles
thermo_dataset = descent.targets.thermo.create_dataset(dhmix_property, ...)
mapped_smiles = descent.targets.thermo.extract_smiles(thermo_dataset)
graphs = ... # convert to Graph

for epoch in range(n_epochs):
    for gas_batch in gas_dataloader:
        # gas-phase stuff ...
        # you could potentially throw to smee here if you can include nonbonded energies
        # just going to assume iterating over gas_batches yields graphs
        predicted_params = [model(graph) for graph in gas_batch]

        ff, topos = convert_espaloma(list(gas_batch), predicted_params, charges=...) # could pre-compute charges by predicting with espaloma or just look up NAGL charges

        # could either compute energy the current way or potentially use smee
        for topo, params, graph in zip(topos, predicted_params, gas_batch):
            coords  = graph.pos #?units?
            E_pred = smee.compute_energy(topo, ff, coords)
            F_pred = -torch.autograd.grad(E_pred.sum(), coords, create_graph=True)[0]
        loss_gas = ...

    if not epoch % 100: # probably too slow to run every epoch        
        params = [model(graph) for graph in graphs]
        charges = ... # predict or cache
        ff_thermo, topos = convert_espaloma(graphs, params, charges=charges)
        topologies_dict = dict(zip(mapped_smiles, topos))
    
        # note the descent predict() function actually runs simulations
        # or reads them from a cached directory, either way
        # the defaults in descent are not very good, 256 mol boxes and very short equilibration/production times.
        # this is where dimsim has the most value in replacing and abstracting compute!
        y_ref, _, y_pred, _ = descent.targets.thermo.predict(
            thermo_dataset, ff_thermo, topologies_dict, output_dir, cached_dir
        )
        loss_thermo = ...

but this is super untested obviously.

[lilyminium]

just hardcoded this assumption that it's called epsilon and sigma