Update! Checkout nequix, a universal interatomic potential with JAX and PyTorch backends.

nequip-eqx

nequip-eqx is a JAX implementation of the neural network interatomic potential NequIP, introduced by Batzner et al. in E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.

The goal of this repository is to offer a simple (<1000 lines of code) implementation while providing competitive performance to existing implementations.

Usage

Installation

pip install nequip-eqx

Training

Models are trained with the nequip_eqx_train command using a single .yml configuration file:

nequip_eqx_train <config>.yml

See configs/3bpa.yml for an example configuration file for training on the 3BPA dataset. Pretrained weights for 3BPA are available in the models/ directory.

tar -C data -xf data/dataset_3BPA.tar.gz  # decompress data
nequip_eqx_train configs/3bpa.yml # takes ~20 hrs on an NVIDIA RTX A5500

Testing

Models can be evaluated with the nequip_eqx_test command by supplying a path to a pretrained model, and a test .xyz file, e.g.:

nequip_eqx_test \
    --model models/nequip_3bpa.eqx \
    --file data/dataset_3BPA/test_300K.xyz

ASE calculator

Using nequip_eqx.calculator.NequipCalculator, you can perform calculations in ASE with a pre-trained NequIP model.

import ase.io
from ase.md.langevin import Langevin
from ase import units

from nequip_eqx.calculator import NequipCalculator

atoms = ase.io.read("data/dataset_3BPA/test_300K.xyz", index=0, format="extxyz")
atoms.calc = NequipCalculator("models/nequip_3bpa.eqx")

dyn = Langevin(
    atoms,
    timestep=0.5 * units.fs,
    temperature_K=300,
    friction=0.01,
    trajectory="md.traj",
)
dyn.run(steps=1000)

Comparison with other codes

In order to verify correctness of the implementation, we compare performance on the 3BPA dataset to two different PyTorch NequIP implementations:

1. The results from Musaelian et al. in "Learning local equivariant representations for large-scale atomistic dynamics", using a version of their nequip repository.
2. The results from Batatia et al. in "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields", using a version of their mace repository.

We use the same hyperparameters as [1], which can be viewed in configs/3bpa.yml, with the following exceptions:

• Instead of hidden irreps of 64x0e + 64x0o + 64x1o + 64x1e + 64x2e + 64x2o + 64x3o + 64x3e, we opt for 128x0e + 128x1o + 128x2e + 128x3o for simplicity with the same feature dimensions.
• It is not clear what initialization was originally used for the radial MLP, but we use Kaiming normal, i.e. sampling from $\mathcal{N}(0, \mathrm{std})$ with $\mathrm{std} = \sqrt{\frac{4.0}{\mathrm{fan\_in}}}$.
• Isolated atom energies are added to each predicted node energy.

Resulting energy (E) and force (F) RMSE in meV and meV/Å respectively.

Code	nequip	mace	nequip-eqx (this repo)
300 K E	3.3 (0.1)	3.1 (0.1)	2.9
300 K F	10.8 (0.2)	11.3 (0.2)	9.5
600 K E	11.2 (0.1)	11.3 (0.3)	10.8
600 K F	26.4 (0.1)	27.3 (0.3)	24.2
1200 K E	38.5 (1.6)	40.8 (1.3)	34.1
1200 K F	76.2 (1.1)	86.4 (1.5)	75.4

Caveats

JAX, by default, uses TensorFloat-32 (tf32) for matmul operations on hardware that supports it, i.e. NVIDIA Ampere GPUs and later. This is typically significantly faster than float32, however is not supported on older GPUs, e.g. V100. For this reason, we force matmuls to use float32 on all hardware. If you are training/running inference exclusively on GPUs that support tf32, this line can be removed.

See also

• mariogeiger/nequip-jax: Another basic implementation of a NequIP style model in JAX.
• e3nn/e3nn-jax: JAX library for E(3)-equivariant neural networks used in this repo.
• patrick-kidger/equinox: JAX library for building neural network architectures, used in this repo.

Citations

@article{batzner20223,
  title={E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials},
  author={Batzner, Simon and Musaelian, Albert and Sun, Lixin and Geiger, Mario and Mailoa, Jonathan P and Kornbluth, Mordechai and Molinari, Nicola and Smidt, Tess E and Kozinsky, Boris},
  journal={Nature communications},
  volume={13},
  number={1},
  pages={2453},
  year={2022},
  publisher={Nature Publishing Group UK London}
}

@article{musaelian2023learning,
  title={Learning local equivariant representations for large-scale atomistic dynamics},
  author={Musaelian, Albert and Batzner, Simon and Johansson, Anders and Sun, Lixin and Owen, Cameron J and Kornbluth, Mordechai and Kozinsky, Boris},
  journal={Nature Communications},
  volume={14},
  number={1},
  pages={579},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

@article{batatia2022mace,
  title={MACE: Higher order equivariant message passing neural networks for fast and accurate force fields},
  author={Batatia, Ilyes and Kovacs, David P and Simm, Gregor and Ortner, Christoph and Cs{\'a}nyi, G{\'a}bor},
  journal={Advances in neural information processing systems},
  volume={35},
  pages={11423--11436},
  year={2022}
}