Struggling to get correlation vectors for structures

[mstapelberg]

Hello!

EDIT:
Turns out I was using the relaxed structures, when I should have used the initial structures when creating the entries. Hope this helps someone else!

Thanks again.

I am trying to identify supercells for a 58V-3Cr-3Ti system (64 atoms). I have used a fine-tuned CHGNet model to relax randomly ordered 64 atom supercells, but am struggling to generate correlation vectors for each structure when following the creating-a-ce.ipynb

I was wondering if there are any suggestions/debug strategies when trying to calculate correlation vectors? New to CE and smol so not too certain what's the best path forward.

The exact error is:

{
	"name": "StructureMatchError",
	"message": "Supercell could not be found from structure",
	"stack": "---------------------------------------------------------------------------
StructureMatchError                       Traceback (most recent call last)
Cell In[19], line 2
      1 structure = lno_entries[1].structure
----> 2 corr = subspace.corr_from_structure(structure)
      4 print(f'The correlation vector for a structure with'
      5       f' composition {structure.composition} is: '
      6       f'\
{corr}')

File ~/.mambaforge/envs/chgnet/lib/python3.10/site-packages/smol/cofe/space/clusterspace.py:684, in ClusterSubspace.corr_from_structure(self, structure, normalized, scmatrix, site_mapping)
    652 \"\"\"Get correlation vector for structure.
    653 
    654 Returns the correlation vector for a given structure. To do this, the
   (...)
    681     array: correlation vector for given structure
    682 \"\"\"
    683 if scmatrix is None:
--> 684     scmatrix = self.scmatrix_from_structure(structure)
    686 occu = self.occupancy_from_structure(
    687     structure, scmatrix=scmatrix, site_mapping=site_mapping, encode=True
    688 )
    689 indices = self.get_orbit_indices(scmatrix)

File ~/.mambaforge/envs/chgnet/lib/python3.10/site-packages/smol/cofe/space/clusterspace.py:835, in ClusterSubspace.scmatrix_from_structure(self, structure)
    833 scmatrix = self._sc_matcher.get_supercell_matrix(structure, self.structure)
    834 if scmatrix is None:
--> 835     raise StructureMatchError(\"Supercell could not be found from structure\")
    836 if np.linalg.det(scmatrix) < 0:
    837     scmatrix *= -1

StructureMatchError: Supercell could not be found from structure"
}

To address this, I tried implementing my own structure matcher with a larger stol, ltol, and angle,

# creating the structure matcher 
from pymatgen.analysis.structure_matcher import StructureMatcher
sm = StructureMatcher(ltol = 5, stol = 5, angle_tol=15,supercell_size='num_sites')

And then when creating the ClusterSubspace:

# define the structure matcher 

subspace = ClusterSubspace.from_cutoffs(
    lno_prim,
    cutoffs={2: 5.5, 3: 4.5}, # will include orbits of 2 and 3 sites.
    basis='sinusoid', # sets the site basis type, default is indicator
    supercell_size='num_sites',
    use_concentration=True,
    num_threads=12,
    supercell_matcher=sm

)

And finally calculating the correlation vector:

structure = lno_entries[1].structure
corr = subspace.corr_from_structure(structure)

print(f'The correlation vector for a structure with'
      f' composition {structure.composition} is: '
      f'\n{corr}')

I get the following error:

{
	"name": "ValueError",
	"message": "get_supercell_matrix cannot be used with the primitive cell option",
	"stack": "---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[17], line 2
      1 structure = lno_entries[1].structure
----> 2 corr = subspace.corr_from_structure(structure)
      4 print(f'The correlation vector for a structure with'
      5       f' composition {structure.composition} is: '
      6       f'\
{corr}')

File ~/.mambaforge/envs/chgnet/lib/python3.10/site-packages/smol/cofe/space/clusterspace.py:684, in ClusterSubspace.corr_from_structure(self, structure, normalized, scmatrix, site_mapping)
    652 \"\"\"Get correlation vector for structure.
    653 
    654 Returns the correlation vector for a given structure. To do this, the
   (...)
    681     array: correlation vector for given structure
    682 \"\"\"
    683 if scmatrix is None:
--> 684     scmatrix = self.scmatrix_from_structure(structure)
    686 occu = self.occupancy_from_structure(
    687     structure, scmatrix=scmatrix, site_mapping=site_mapping, encode=True
    688 )
    689 indices = self.get_orbit_indices(scmatrix)

File ~/.mambaforge/envs/chgnet/lib/python3.10/site-packages/smol/cofe/space/clusterspace.py:833, in ClusterSubspace.scmatrix_from_structure(self, structure)
    820 def scmatrix_from_structure(self, structure):
    821     \"\"\"Get supercell matrix from a given structure.
    822 
    823     Obtain the supercell structure matrix to convert the prim structure
   (...)
    831         ndarray: matrix relating given structure and prim structure.
    832     \"\"\"
--> 833     scmatrix = self._sc_matcher.get_supercell_matrix(structure, self.structure)
    834     if scmatrix is None:
    835         raise StructureMatchError(\"Supercell could not be found from structure\")

File ~/.mambaforge/envs/chgnet/lib/python3.10/site-packages/pymatgen/analysis/structure_matcher.py:1072, in StructureMatcher.get_supercell_matrix(self, supercell, struct)
   1066 \"\"\"
   1067 Returns the matrix for transforming struct to supercell. This
   1068 can be used for very distorted 'supercells' where the primitive cell
   1069 is impossible to find.
   1070 \"\"\"
   1071 if self._primitive_cell:
-> 1072     raise ValueError(\"get_supercell_matrix cannot be used with the primitive cell option\")
   1073 struct, supercell, fu, s1_supercell = self._preprocess(struct, supercell, niggli=False)
   1075 if not s1_supercell:

ValueError: get_supercell_matrix cannot be used with the primitive cell option"
}

Thank you and happy to provide further information!

Myles

[lbluque]

Hi @mstapelberg 👋

Glad you found a workaround for your problem!

Though it sounds like your relaxed structures have shifted far (maybe too far?) from the initial structure. In some of these cases the correlations for the initial structure and relaxed structure may no longer be the same.

For such cases with large relaxations what you tried to do by raising the tolerance in the StructureMatcher is exactly what I would recommend. To avoid the second error you got, I recommend passing the tolerance keywords (stol, atol, or ltol) directly to the ClusterSubspace constructor, instead of creating the structure matcher independently. This will take care of creating the correct structure matcher for you, but with the given tolerances.

• In general all the arguments to create the structure matcher can be passed as keywords to the ClusterSubspace.

Hope this helps!

[mstapelberg]

Hi @lbluque ,

Thanks for your reply and suggestion! I saw how to add stol, atol, and ltol based on the LiMnO DRX Cluster Expansion Tutorial.

Thanks for your help!

Best,
Myles