Wrap atom coordinates in ase.Atoms preprocessing

[kyonofx]

In our PBC graph construction algorithm, when a pair of atoms become very far away from each other in terms of unwrapped coordinates (which usually happens in long simulations), they may lose track of each other because only a finite number of supercells are considered during the graph construction process. To resolve this, we wrap the atom coordinates before feeding them into the graph construction algorithm and the neural network model when converting ase.Atoms.

Two test cases under tests/core/preprocessing/test_atoms_to_graphs.py that directly check the output from AtomsToGraphs will need to be updated.

[kyonofx]

@mshuaibii @lbluque Would be great if you could take a look :)

[mshuaibii]

Looks good. Let's update the tests to get them to pass. Rather than replacing the source of truth we can actually change this line -

fairchem/tests/core/preprocessing/test_atoms_to_graphs.py

Line 113 in 2a2a3df

np.testing.assert_allclose(act_positions, positions)

Rather than verifying the positions are identical, we should be able to compute a distance accounting for pbc to verify they are correct -

fairchem/src/fairchem/core/modules/evaluator.py

Line 272 in 029d4d3

def min_diff(

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Files	Coverage Δ
src/fairchem/core/preprocessing/atoms_to_graphs.py	88.42% <100.00%> (+0.81%)	⬆️

[mshuaibii]

LGTM!