computing energy of a benzene crystal results in a RuntimeError about ase_ani HOT 7 CLOSED

Please note that ANI was not trained with PBC, so do not expect any quantitative accuracy at this point.
AFAIK, benzene crystal is not orthorhombic (angles =/=90deg ). I think we don't have this functionality in the public repo. Only angles = 90 are supported now. Feel free to contact us if you are interested in this functionality.

from ase_ani.

I understand ANI has not been trained to energies from periodic DFT and these are really just tests of the implementation not accuracy. This particular cell is orthorhombic:

_refine_ls_R_factor_gt 0.053
_refine_ls_wR_factor_gt 0.053
_diffrn_radiation_probe x-ray
_symmetry_cell_setting orthorhombic
_symmetry_space_group_name_H-M 'P b c a'
symmetry_Int_Tables_number 61
loop
_symmetry_equiv_pos_site_id
_symmetry_equiv_pos_as_xyz
1 x,y,z
2 1/2-x,-y,1/2+z
3 -x,1/2+y,1/2-z
4 1/2+x,1/2-y,-z
5 -x,-y,-z
6 1/2+x,y,1/2-z
7 x,1/2-y,1/2+z
8 1/2-x,1/2+y,z
_cell_length_a 7.287(6)
_cell_length_b 9.20(2)
_cell_length_c 6.688(9)
_cell_angle_alpha 90
_cell_angle_beta 90
_cell_angle_gamma 90
_cell_volume 448.366
_exptl_crystal_colour colorless
cell_formula_units_Z 4
loop
_atom_site_label
_atom_site_type_symbol
_atom_site_fract_x
_atom_site_fract_y
_atom_site_fract_z
C1 C -0.0537(8) 0.1425(9) 0.0097(12)
H1 H -0.085(6) 0.246(7) 0.034(8)
C2 C 0.0840(7) 0.0924(10) 0.1373(10)
H2 H 0.140(6) 0.156(6) 0.219(8)
C3 C -0.1343(7) 0.0521(9) -0.1235(12)
H3 H -0.220(6) 0.080(6) -0.204(9)
C1D C 0.0537 -0.1425 -0.0097
H1D H 0.085 -0.246 -0.034
C2D C -0.084 -0.0924 -0.1373
H2D H -0.14 -0.156 -0.219
C3D C 0.1343 -0.0521 0.1235
H3D H 0.22 -0.08 0.204

from ase_ani.

If you have an input that can be read by ASE that you have an energy and force to reproduce I'd be happy to test on my machine. Thanks for any help you can provide.

from ase_ani.

Hi,

1. the current code only support rectangular cells.
2. if the edge length of your cell is < 2x the cutoff the code will crash with a memory error (same as your error)
3. as Oles mentioned, these potentials were not trained to work on organic crystals, so I wouldn’t expect this to work well.
4. can you define ‘the energy drifts?”. As in the energy is not conserving in NVE? If so, what precision are you running torchani at? You might want to see if switching to double helps.

from ase_ani.

Thank you for the quick responses. I sounds like issue 2 that Justin points out is a possibility as the sidelengths of this cell are ~ 7 A whereas I believe the cutoff is ~5.5 A. I'll test with a supercell. 4. Sure, the total energy in an NVE dynamics with 0.5 fs time step is not conserved (or even close really) over the coarse of a 5 ps trajectory. Torchani is being run in double precision and the energy agrees to fairly high precision with our in house implementation which is also using double precision.

from ase_ani.

Interesting, I’ve thoroughly tested conservation in neurochem (ASE_ANI) for gas phase systems. I also tested conservation heavily for periodic systems in recent work (https://arxiv.org/abs/1811.01914). There is some level of drift but what we saw made sense given time scale, time step, and precision of the model itself. How much drift are you seeing? Can you quantify?

from ase_ani.

I tested a larger cubic cell of water with side length 12 A. ANI_ASE now runs without error. Tentatively the energy conservation looks better with this implementation. The disagreement in energy between this implementation and torchani is about 1.0e-4 a.u. for a single point energy evaluation on this system so something seems inconsistent. Feel free to close this issue.

from ase_ani.