Update BAWL to use reduced composition and allow for primitive structure reduction.

[msiron-entalpic]

Problem: currently for some structures, the conventional cell returns a different hash than the primitive cell. For supercell, the full composition is used, so also a different hash is returned. (pointed out by @bkmi in #20 )

Solution: Ideally a supercell, or a conventional vs. primitive cell of a material should return the same hash. Propose update to BAWL to account for both reduced composition (which we initially did in LeMat-Bulk) and to account for transformation to primitive cell using Moyopy.

To do: would welcome benchmarks that transform materials to either primitive or conventional to test whether various hash function well against that. Or move to switch all hash to consider only the primitive transformation via Moyo?

In PR it is not reduced to primitive by default, it must be turned on:
bh = BAWLHasher(primitive_reduction=True)

[Ramlaoui]

As suggested by @msiron-entalpic, we should enable this for any hasher or similarity matcher for fairness in the benchmark and enable them altogether when we want to reduce to the most primitive cell cf #22.

[bkmi]

There is small disagreement between moyo and spglib in 2% of the MP structures' space group. Is it obvious to you all what is the source of truth?
https://github.com/spglib/moyo/blob/main/bench/mp/analysis.ipynb

[bkmi]

The hasher worked trivially with

BAWLHasher(primitive_reduction=True).get_material_hash(s.make_supercell((2,1,1))) == BAWLHasher(primitive_reduction=True).get_material_hash(s)

[msiron-entalpic]

@bkmi spacegroup determination in Materials Project uses spglib under the hood as built into Pymatgen with a symmetry precision of 0.01 and an angle tolerance of 5 by default. By default, spglib recommends a symmetry precision of 1e-5 and a angle tolerance of -1. moyo is the newer library from the same group which does spglib. Either way, symmetry detection is not always unique and depends on precision parameters chosen, we are simply aiming for speed! Hope that helps!!

[bkmi]

Do you know the default parameters for moyo?

[bkmi]

btw these parameters are quite strict. I might suggest something more premissive, such as symmetry precision = 0.1 (or 0.01 like pymatgen). Maybe also the angle default from pymatgen.

[bkmi]

More experiments, unfortunately all is not well.

I am using a version where I allow for a higher tolerance for matches...
https://github.com/bkmi/lematerial-hasher/tree/fix/bawl-primitive

but this causes hanging behavior. I think it's due to the underlying moyo library.

from material_hasher.hasher.bawl import BAWLHasher
from pymatgen.io.cif import CifParser
from pymatgen.transformations.standard_transformations import PerturbStructureTransformation
from pymatgen.io.ase import AseAtomsAdaptor
from copy import deepcopy

parser = CifParser("the.cif")
s = parser.get_structures()[0]
sprc = deepcopy(s).make_supercell((2,1,1))

BAWLHasher(primitive_reduction=False).get_material_hash(sprc) == BAWLHasher(primitive_reduction=False).get_material_hash(s)
# false 👍 

BAWLHasher(primitive_reduction=True).get_material_hash(sprc) == BAWLHasher(primitive_reduction=True).get_material_hash(s)
# true 👍 

atoms = AseAtomsAdaptor.get_atoms(deepcopy(sprc))
atoms.rattle(0.01)
rsprc = AseAtomsAdaptor.get_structure(atoms)
BAWLHasher(primitive_reduction=True).get_material_hash(rsprc) == BAWLHasher(primitive_reduction=True).get_material_hash(s)
# this hangs 👎 

the.cif

# generated using pymatgen
data_Fe2O3
_symmetry_space_group_name_H-M   'P 1'
_cell_length_a   5.13463532
_cell_length_b   5.13463532
_cell_length_c   5.50540479
_cell_angle_alpha   62.20387762
_cell_angle_beta   62.20387762
_cell_angle_gamma   60.00000026
_symmetry_Int_Tables_number   1
_chemical_formula_structural   Fe2O3
_chemical_formula_sum   'Fe4 O6'
_cell_volume   105.92154660
_cell_formula_units_Z   2
loop_
 _symmetry_equiv_pos_site_id
 _symmetry_equiv_pos_as_xyz
  1  '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
  Fe  Fe0  1  0.85364782  0.85364782  0.43905654  1
  Fe  Fe1  1  0.35364782  0.35364782  0.93905654  1
  Fe  Fe2  1  0.14635218  0.14635218  0.56094346  1
  Fe  Fe3  1  0.64635218  0.64635218  0.06094346  1
  O  O4  1  0.44604828  0.05395172  0.75000000  1
  O  O5  1  0.05395172  0.75000000  0.75000000  1
  O  O6  1  0.75000000  0.44604828  0.75000000  1
  O  O7  1  0.94604828  0.25000000  0.25000000  1
  O  O8  1  0.25000000  0.55395172  0.25000000  1
  O  O9  1  0.55395172  0.94604828  0.25000000  1

I brought this up in spglib/moyo#89

[msiron-entalpic]

Do you know the default parameters for moyo?

I believe it is 1e-4 for symprec and for angle_tolerance it should be the same as spglib

[msiron-entalpic]

btw these parameters are quite strict. I might suggest something more premissive, such as symmetry precision = 0.1 (or 0.01 like pymatgen). Maybe also the angle default from pymatgen.

Unsure about the implications of this, but let me make a fix to be able to pass on custom tolerance parameters for BAWL!

[bkmi]

I did that in my fork. Sorry I didn't pull request it. I'll do it now.

[bkmi]

Done

[Ramlaoui]

LGTM

[bkmi]

I used relatively permissive parameters. I think we should try to figure out what's the deal with the hanging before merging this (using the permissive parameters).

I think making it work with permissive parameters is important.

[msiron-entalpic]

I think making it work with permissive parameters is important.

Actually you're right, I was thinking of downstream effect since we use this code for labeling fingerprints for LeMat-Bulk, but currently we use Pymatgen with the less stringent symprec, so if we move to Moyo we should also use the less stringent parameters by default.

[bkmi]

It sounds like the hanging might have been caused by using degrees when I should have used radians.

I'm busy for a while but will follow up later. This tool would be extremely useful for us.

[bkmi]

I fixed it.

Specifically, I finished my pull request onto fix/bawl-primitive. Namely, this one #23. Please merge my fork into into fix/bawl-primitive and then merge this into main!

Thank you!!

[bkmi]

The hasher does not work reliably with moyo. I am investigating making pymatgen the default.

It is not recommended to use angle_tolerance in any case... according to this one random post (and the fact that using angle_tolerance lead to some cases hanging). spglib/spglib#567

--

p.s. it remains to be seen whether standard, primitive cells (used in spglib & moyo) are the same as niggli reduced primitive cells. pymatgen claims that standard, primitive are defined in the source

Setyawan, W., & Curtarolo, S. (2010). High-throughput electronic band structure calculations: Challenges and tools. Computational Materials Science, 49(2), 299-312. doi:10.1016/j.commatsci.2010.05.010.

[bkmi]

I cannot comment yet on the quality of the hash, but I'm getting much more stable behavior using my new implementation that uses the reduction and primitive functions from pymatgen.