Regarding the selection of stress units

[Zzz3626]

Dear Community,

I noticed that the docs states that NequIP always expresses stress in the "consistent" units of energy / length^3, which are not the typical physical units used by many codes for stress. Does this mean that I cannot use bar as a unit and need to convert it to eV/ų?

I appreciate any insight and help.

Thank you!

[cw-tan]

Hi @Zzz3626 ,

Yes, if your energies are in eV, and your positions (and cell) are in angstroms. If you decide to use hartree energies or bohr lengths, or any specific combination, the stress has to be consistent with them (same with forces).

[Zzz3626]

Thanks a lot for the reply. In LAMMPS' metal unit system, energy is in eV, force is in eV/Angstrom, but stress is in bars. How can I make my model consistent?

[cw-tan]

We just apply automatic differentiation on the PyTorch model without having a specific preference for units -- that's why the units of the underlying dataset being used to train the model must be self-consistent. For LAMMS's metal units, as long as your energy is in eV and length units are in angstrom (and everything, including stresses, are in consistent units) during training, the stress will be computed internally correctly such that whatever LAMMPS spits out is in expected metal units, i.e. bars. You can sanity check this by running your model in an ASE calculator and through LAMMPS, and comparing that the stresses for the same structure are related by the expected sign and units changes -- see https://github.com/mir-group/pair_nequip/blob/main/tests/test_python_repro.py and relevant changes in the PR https://github.com/mir-group/pair_nequip/pull/52/files