❓ [QUESTION] Dataset preprocessing computation times about nequip HOT 2 CLOSED

Hi @pdobbelaere, I loaded your structures and ran the ASE neighborlist calculation that we use internally. I can confirm that I see the same behavior:

A: PBC + cell are set, the neighborlist is aware of periodic images but since the cell is large never sees them, all is fine
B: PBC is false, cell is set, much slower.
C: PBC is false, cell is not there, this is equivalent to A since the cutoff is small enough in A that it's the same as if it was just in vacuum

This is the neighborlist we use. Note that the docs say:

Atoms outside periodic boundaries are mapped into the box. Atoms
outside nonperiodic boundaries are included in the neighbor list
but complexity of neighbor list search for those can become n^2.

Looking at your structure your atoms lie right on the border, hence you're seeing the N^2 scaling of ASE. When I shift them to the middle of the cell, the behavior goes away.

import time
import numpy as np

from ase.io import read
from ase.neighborlist import neighbor_list

b = read('./clusters_b.xyz')

s = time.time()
ib, jb = neighbor_list('ij', b, cutoff=5., self_interaction=False)
e = time.time()
print(e-s)

# simple cubic cell
cell = b.get_cell()
center = cell[0][0]/ 2.

# shift to middle of box
p = b.get_positions()
p += np.array([center, center, center])
b.set_positions(p)
b.wrap()

s = time.time()
ib, jb = neighbor_list('ij', b, cutoff=5., self_interaction=False)
e = time.time()
print(e-s)

Gives:

9.93855881690979
0.07891488075256348

from nequip.

Thanks very much for the speedy reply. That is good to know.

from nequip.