Fix grid overlap

[fxcoudert]

This fixes a bug in the current code, where grid points at the edges ($x=1$ or $y=1$ or $z=1$, in fractional coordinates) are included, which leads to duplication (because they are the same as $x=0$, etc).

Given the use of periodic boundary conditions for the voxels, duplication is wrong. It leads to give double weight to points on the faces of the cell, and leads to wrong PBC application once the voxels are processed.

np.linspace() accepts a endpoint=False optional arg, which keeps the total number of points but does not include the endpoint. See in this simple example:

>>> import numpy as np
>>> np.linspace(0, 1, 5)
array([0.  , 0.25, 0.5 , 0.75, 1.  ])
>>> np.linspace(0, 1, 5, endpoint=False)
array([0. , 0.2, 0.4, 0.6, 0.8])

[adosar]

This should not affect the case where cubic_box is set to True (cartesian coords are used in this case), right? Meaning that only the fractional coordinates case should be modified, i.e.:

if cubic_box:
    d = length / 2
    probe_coords = np.linspace(0-d, 0+d, self.grid_size)  # Cartesian, leave it as is.
    self._simulation_box = self.structure
else:
    probe_coords = np.linspace(0, 1, self.grid_size, endpoint=False)  # Fractional.

[fxcoudert]

Hum, I wonder. If you set cubic_box and the box is not cubic, you're definitely biasing the sampling anyway. So maybe it does not matter much if you bias more?

On the other hand, if you set cubic_box but the box is actually cubic then you're actually still giving double weight to those points. That, at least, should be avoided.

I don't have a strong opinion. But I don't think cubic_box is physically meaningful anyway… have you tested whether it improves accuracy of the trained models?

[adosar]

On the other hand, if you set cubic_box but the box is actually cubic then you're actually still giving double weight to those points. That, at least, should be avoided.

Yep, that is correct. To ensure correct PBC, the argument endpoint will be added in both cases.

I don't have a strong opinion. But I don't think cubic_box is physically meaningful anyway… have you tested whether it improves accuracy of the trained models?

I have not tested the performance of models with cubic_box set to True, but I believe it might be beneficial if the dataset under consideration contains many materials with large deviations from orthogonality.