AIMNet2 Fails On Adamantane

[corinwagen]

When I run a PBC calculation on adamantane using the current branch of the code, I get just nan values back.

Here's aimnet2_fail.py:

from ase.io import read
from aimnet2calc import AIMNet2ASE

calc = AIMNet2ASE('aimnet2')
atoms = read("adamantane.extxyz")
atoms.calc = calc
atoms.get_potential_energy()

And here's adamantane.extxyz (taken from the SI of the r2SCAN-3c paper):

26
Lattice="264.58862450 0.00000000 0.00000000 0.00000000 264.58862450 0.00000000 0.00000000 0.00000000 264.58862450" Properties=species:S:1:pos:R:3
c        0.00000102      -0.00000242       1.77273919
c        0.63246610     262.93251026     264.58869533
c      263.95616059       1.65611486     264.58869514
c      262.93250743     263.95616524      -0.00006534
c        1.14670307     264.07566999     263.70036917
c        1.65611690       0.63245867      -0.00006291
c      263.44192114       0.51295539     263.70036931
c      264.07566759     263.44191859       0.88825074
c        0.00000019      -0.00000065     262.81588362
c        0.51295396       1.14670273       0.88825103
h      262.10014570     264.30751367       0.62510959
h      263.13942217       2.04214911       0.62538716
h      262.54646928     263.13941971     263.96323918
h        0.28111555     262.10014031     263.96351644
h        2.48847747       0.28111351       0.62510826
h      264.30751053       2.48848158     263.96351417
h        1.96537171     263.70941606     263.06626904
h        0.80706748     264.22754426       2.42492481
h      262.62325182       0.87920706     263.06626927
h      263.70941681     262.62325285       1.52235358
h      264.22754565     263.78155762     262.16369529
h        2.04215132       1.44920559     263.96323582
h        0.36108185       0.80706968     262.16369848
h      263.78155514       0.36108089       2.42492490
h        0.87920761       1.96537186       1.52235248
h        1.44920588     262.54647754       0.62538560

This is what happens when I run the script:

Agincourt:peregrine cwagen$ python aimnet2_fail.py
/Users/cwagen/business_code/peregrine/subrepos/aimnet2calc/aimnet2calc/__init__.py:22: UserWarning: PySisiphus is not installed. AIMNet2Pysis will not be available.
  warnings.warn("PySisiphus is not installed. AIMNet2Pysis will not be available.")
Downloading model file from https://github.com/zubatyuk/aimnet-model-zoo/raw/main/aimnet2/aimnet2_wb97m_0.jpt
Saved to  /Users/cwagen/business_code/peregrine/subrepos/aimnet2calc/aimnet2calc/assets/aimnet2/aimnet2_wb97m_0.jpt
Atoms(symbols='C10H16', pbc=True, cell=[264.5886245, 264.5886245, 264.5886245], calculator=AIMNet2ASE(...))
Switching to DSF Coulomb for PBC
OMP: Info #276: omp_set_nested routine deprecated, please use omp_set_max_active_levels instead.
{'charges': array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan,
       nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan],
      dtype=float32), 'energy': array([nan])}

I'm not sure what the OMP error means; possibly something from numba? cf. numba/numba#5520. AIMNet2's worked on other periodic systems for me, so not sure what's going wrong here.

[husakm]

What uintis do yo use ?
264.58862450 as lattice parameter in Ångström (dafault AS units, you do not change them) is for sure wrong.
The position do not give sense as well ...

[corinwagen]

The structure has the adamantane molecule at the unit-cell boundary, and uses an extremely large unit cell to model a single gas-phase molecule. There's a corresponding solid-phase crystal structure too, and the eval uses the difference between these two structures to compute the lattice energy.

Every molecule in the X23 set is like this; most of them work!

[husakm]

The description of the molecule in .xyz file is wrong. Try to load it in any visualization program (e.g. Avogadro). The molecule explode in small parts 256 A away not connected together.
Even if you use slab to model gasss phase, you use about 10 A of space to mimic vacuum , not 264 A.

[corinwagen]

It looks fine in the ASE visualizer:

Try running this in a Jupyter notebook:

from ase.visualize import view
from ase.io import read
atoms = read("adamantane.extxyz")
view(atoms, repeat=(2,2,2))

Is it possible that the lattice information isn't getting read by Avogadro? I've had this issue with .extxyz files before.

I agree that 264 Å seems like overkill for mimicking vacuum; I'm just following literature precedent here. Again, this works for most NNPs, and other molecules with 264 Å buffer work with AIMNet2.