`CrystalNN` gives different results for fractional vs. Cartesian coordinates

[nayoung10]

Python version

Python 3.10.12

Pymatgen version

2024.10.22

Operating system version

Ubuntu 20.04.6 LTS

Current behavior

I noticed that CrystalNN behaves differently depending on whether the input structure is defined using fractional or Cartesian coordinates.

For example, consider the following two structures, which are essentially equivalent (their RMSD from StructureMatcher is 1.84e-07):

structure_cart = Structure(
    coords=data['gt_coords'],
    species=data['atom_types'],
    lattice=data['cell'],
    coords_are_cartesian=True,
)
structure_frac = Structure(
    coords=cart2frac(data['gt_coords'], data['cell']) % 1.0,
    species=data['atom_types'],
    lattice=data['cell'],
    coords_are_cartesian=False,
)

# Get subgraphs
crystnn = CrystalNN()

struct_graph_cart = StructureGraph.from_local_env_strategy(structure=structure_cart, strategy=crystnn)
molecules_cart = struct_graph_cart.get_subgraphs_as_molecules()
len(molecules_cart) # 2

struct_graph_frac = StructureGraph.from_local_env_strategy(structure=structure_frac, strategy=crystnn)
molecules_frac = struct_graph.get_subgraphs_as_molecules()
len(molecules_frac) # 0

Despite the two structures being identical, the number of molecules extracted differs.

This behavior is important since tools such as MOFChecker rely on get_subgraphs_as_molecules to determine whether a structure contains isolated molecules. In my case, the same structure passes the test when using fractional coordinates but fails when using Cartesian coordinates.

Could you kindly look into this issue? I have attached the relevant data file below for reproducibility.

Thank you very much for your time and support!

Expected Behavior

They should behave the same.

Minimal example

Please see above.

Relevant files to reproduce this bug

data.pt.zip

[esoteric-ephemera]

Hey @nayoung10, to isolate what the issue could be, can you try:

structure_frac = Structure(
    coords=structure_cart.frac_coords,
    species=data['atom_types'],
    lattice=data['cell'],
    coords_are_cartesian=False,
).to_primitive()

structure_cart = structure_cart.to_primitive()

The idea is to see if you're looking at different representations of the same cell and therefore getting different connectivity patterns because there are simply different numbers of atoms in the cell

Also I'm not sure what cart2frac is doing - is this from another package?

[nayoung10]

Hi @esoteric-ephemera,

cart2frac is simply a function that converts Cartesian coordinates into fractional coordinates as follows:

def cart2frac(coord, cell):
    # pylint: disable=E1102
    invcell = torch.linalg.inv(cell)
    return coord @ invcell

I have tried your code and I get the same number of atoms in the cell

struct_cart_primitive = struct_cart.get_primitive_structure()
struct_frac_primitive = struct_frac.get_primitive_structure()

struct_cart_primitive.frac_coords.shape, struct_frac_primitive.frac_coords.shape # both (156,3)

It really seems like the results differ depending on whether fractional or Cartesian coordinates are used. Could you please look into this issue?

Thank you very much for your time and support!

[esoteric-ephemera]

It seems more likely that the conversion from numpy --> torch --> numpy would result in some precision loss that affects these lower symmetry structures, since I can't reproduce this for higher symmetry stuff. Can you use just the pymatgen-constructed frac_coords to test this? Or swap torch.linalg.inv for numpy.linalg.inv?