Nequix

Source code for the Nequix foundation model, and Phonon fine-tuning (PFT).

Usage

Installation

pip install nequix

or for torch

pip install nequix[torch]

ASE calculator

Using nequix.calculator.NequixCalculator, you can perform calculations in ASE with a pre-trained Nequix model.

from nequix.calculator import NequixCalculator

atoms = ...
atoms.calc = NequixCalculator("nequix-mp-1", backend="jax")

or if you want to use the faster PyTorch + kernels backend

...
atoms.calc = NequixCalculator("nequix-mp-1", backend="torch")
...

NequixCalculator

Arguments

• model_name (str, default "nequix-mp-1"): Pretrained model alias to load or download.
• model_path (str | Path, optional): Path to local checkpoint; overrides model_name.
• backend ({"jax", "torch"}, default "jax"): Compute backend.
• capacity_multiplier (float, default 1.1): JAX-only; padding factor to limit recompiles.
• use_compile (bool, default True): Torch-only; on GPU, uses torch.compile().
• use_kernel (bool, default True): Torch-only; on GPU, use OpenEquivariance kernels.

Training

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

nequix_train <config>.yml

or for Torch

# Single GPU
uv sync --extra torch
uv run nequix/torch/train.py <config>.yml
# Multi-GPU
uv run torchrun --nproc_per_node=<gpus> nequix/torch/train.py <config>.yml

To reproduce the training of Nequix-MP-1, first clone the repo and sync the environment:

git clone https://github.com/atomicarchitects/nequix.git
cd nequix
uv sync

Then download the MPtrj data from https://figshare.com/files/43302033 into data/ then run the following to extract the data:

bash data/download_mptrj.sh

Preprocess the data into .aselmdb files:

uv run scripts/preprocess_data.py data/mptrj-gga-ggapu data/mptrj-aselmdb

Then start the training run:

nequix_train configs/nequix-mp-1.yml

This will take less than 125 hours on a single 4 x A100 node (<25 hours using the torch + kernels backend). The batch_size in the config is per-device, so you should be able to run this on any number of GPUs (although hyperparameters like learning rate are often sensitive to global batch size, so keep in mind).

Phonon Fine-tuning (PFT)

First sync extra dependencies with

uv sync --extra pft

Phonon Calculations

We provide pretrained model weights for the co-trained (better alignment with MPtrj) and non co-trained models in models/nequix-mp-1-pft.nqx and nequix-mp-1-pft-nocotrain.nqx respectively. See nequix-examples for examples on how to use these models for phonon calculations with both finite displacement, and analytical Hessians.

Training

Data for the PBE MDR phonon database was originally downloaded and preprocessed with:

bash data/download_pbe_mdr.sh
uv run data/split_pbe_mdr.py
uv run scripts/preprocess_data_phonopy.py data/pbe-mdr/train data/pbe-mdr/train-aselmdb
uv run scripts/preprocess_data_phonopy.py data/pbe-mdr/val data/pbe-mdr/val-aselmdb

However we provide preprocessed data which can be downloaded with

bash data/download_pbe_mdr_preprocessed.sh

To run PFT without co-training run:

uv run nequix/pft/train.py configs/nequix-mp-1-pft-no-cotrain.yml

To run PFT with co-training run (note this requires mptrj-aselmdb preprocessed):

uv run nequix/pft/train.py configs/nequix-mp-1-pft.yml

Citation

@article{koker2026pft,
  title={{PFT}: Phonon Fine-tuning for Machine Learned Interatomic Potentials},
  author={Koker, Teddy and Gangan, Abhijeet and Kotak, Mit and Marian, Jaime and Smidt, Tess},
  journal={arXiv preprint arXiv:2601.07742},
  year={2026}
}

@article{koker2025training,
  title={Training a foundation model for materials on a budget},
  author={Koker, Teddy and Kotak, Mit and Smidt, Tess},
  journal={arXiv preprint arXiv:2508.16067},
  year={2025}
}