Skewed crystal cells #2
Description
Dear Authors,
Congrats on the great publication and result!
I have had a look at the shared dataset and spotted a potential issue with ~20% of the supercells used. I don't think it invalidates your wonderful results, but, possibly, filtering those out may even improve them due to cleaner data.
In the paper, you state that you build "a supercell of the structure to ensure that each dimension is larger than 9 Å". However, looking at the shared universal/lattice.npy file, I notice that minimal inter-planar distance for the cell parallelepiped (and hence the smallest potential distance between an atom and its self-images) is sometimes much smaller than 9A, going down to nearly 2A (e.g., for mp_753092).
I assume this was caused by building the supercell with diagonal transformation (formula 1 in the paper) even for cells with very acute/obtuse angles between the lattice vectors. I think, a better way would be to find an arbitrary integer-valued lattice transformation (new_lattice = M.dot(old_lattice), where M is integer and has positive non-zero determinant), such that the parallelepiped formed by the new lattice can fit a sphere of some minimal diameter (e.g., 9A). You can check the code I've been using in my MD simulations which does precisely that: https://github.com/SiLiKhon/supercell_finder/.
Hope this makes sense.
Best,
Artem