Changes to predicted energy for same model and structure

[ElliottKasoar]

Describe the bug

The predicted energy returned when using the same model and structure appears to have changed.

To Reproduce

from alignn.ff.ff import AlignnAtomwiseCalculator
from ase.io import read

struct = read("tests/data/NaCl.cif")
struct.calc = AlignnAtomwiseCalculator(path="tests/models/v5.27.2024")
print(struct.get_potential_energy())

where NaCl.cif is:

data_image0
_chemical_formula_structural       NaClNaClNaClNaCl
_chemical_formula_sum              "Na4 Cl4"
_cell_length_a       5.64
_cell_length_b       5.64
_cell_length_c       5.64
_cell_angle_alpha    90.0
_cell_angle_beta     90.0
_cell_angle_gamma    90.0

_space_group_name_H-M_alt    "P 1"
_space_group_IT_number       1

loop_
  _space_group_symop_operation_xyz
  'x, y, z'

loop_
  _atom_site_type_symbol
  _atom_site_label
  _atom_site_symmetry_multiplicity
  _atom_site_fract_x
  _atom_site_fract_y
  _atom_site_fract_z
  _atom_site_occupancy
  Na  Na1       1.0  0.0  0.0  0.0  1.0000
  Cl  Cl1       1.0  0.5  0.0  0.0  1.0000
  Na  Na2       1.0  0.0  0.5  0.5  1.0000
  Cl  Cl2       1.0  0.5  0.5  0.5  1.0000
  Na  Na3       1.0  0.5  0.0  0.5  1.0000
  Cl  Cl3       1.0  0.0  0.0  0.5  1.0000
  Na  Na4       1.0  0.5  0.5  0.0  1.0000
  Cl  Cl4       1.0  0.0  0.5  0.0  1.0000

and v5.27.2024 contains the downloaded best_model.pt and config.json files.

This produces different results with different versions of alignn: 2024.05.27 and 2024.8.30 return -11.148092269897461, whereas 2024.10.30 and 2024.12.02 return 143.85162.

Expected behavior

My expectation was that, for the same model and structure, the value returned wouldn't change. Perhaps this relates to:

Faster ALIGNN-FF pretrained model with radius cutoff

Desktop (please complete the following information):

• OS: MacOS 15.2, and Ubuntu 24.04.1