Converting units of trained Allegro/NequIP model to match LAMMPS

[ankur56]

I have trained a model using Allegro and am now using it to perform simulations on a system with Lammps. However, I have noticed a discrepancy in the units used by both programs. I chose to use Angstroms for positions, Hartrees for energies, and Hartrees/Angstroms for forces when training my model with Allegro, but Lammps uses Hartrees and Bohrs for electron-style units.

My question is, should I retrain my model using the same units as Lammps to ensure consistent results across both the trained model and Lammps? If I continue to use my current model, I understand that the energies will be computed in Hartree correctly, but I may not be able to set the force threshold or other internal properties in the correct units.

What are your thoughts and suggestions on this issue? How can I resolve this discrepancy and ensure the accuracy of my simulations?

[Linux-cpp-lisp]

Hi @ankur56 ,

The final step in our default NequIP/Allegro models is RescaleEnergyEtc, which scales the network's predictions into "physical" or "real" units, i.e. those in which it was trained. This rescaling can be easily modified in a deployed model to create a new deployed model that predicts in different energy units. You can use this to create a model that takes Angstroms and predicts eV (LAMMPS metal units) or Angstrom and kcal/mol (real units). Please note that it is not straightforwardly possible to change the length units of a trained model, only the energy units.

Given a deployed model deployed.pth, it should basically be this:

import torch
from nequip.scripts.deploy import load_deployed_model
model, metadata = load_deployed_model("deployed.pth", device="cpu", set_global_options=False, freeze=False)
rescale_output = model.model   # if nequip>=0.6.0
rescale_output = model              # else (older versions)
assert rescale_output.original_name == "RescaleOutput"
with torch.no_grad():
    rescale_output.scale_by *= conversion_factor_training_energy_units_to_desired_energy_units
torch.jit.save(model, "deployed-newunits.pth", _extra_files=metadata)

Definitely check your predictions in LAMMPS after doing this to confirm their sanity.

I would strongly recommend against using a model in LAMMPS whose units are inconsistent with the units LAMMPS is using; that seems like a way for a lot of things to go wrong.

[yangqifanshihaoren]

Then how can I change my stress units? the stress units in vasp(0.1Gpa) and lammps(10-4Gpa) are different.

[yangqifanshihaoren]

Thank you!

[Linux-cpp-lisp]

The units of your model will be whatever units your training dataset is in. The units must also be consistent, as in, the units of other quantities such as stress must be derived from the chosen energy and length units.

[yangqifanshihaoren]

Thank you for replying. Does this means I should convert the training dataset's stress to a new unit corresponding with energy(eV) and length(A)'s units (I chosed in my dataset) before training?
And when I want to use the trained model in lammps for npt simulation, should I change the depolyed model's stress unit to the unit the same as lammps? Or I just use the nequip pair_style and don't have to change anything of the model?