FastCHGNet

This is the official implementation of FastCHGNet, a universal interatomic potential for efficient training and inference.

Installation

FastCHGNet requires Python 3.10 or newer and a C/C++ compiler. Install from the repository root:

pip install -e .

For ASE LMDB datasets, install with pip install -e '.[lmdb]'.

Inference

The pretrained FastCHGNet model predicts energy, forces, stress, and magnetic moments, with forces and stress predicted directly:

from ase.io import read
from fastchgnet.model.dynamics import FastCHGNetCalculator

atoms = read("examples/mp-18767-LiMnO2.cif")
atoms.calc = FastCHGNetCalculator(
    model="fast",  # Non-conservative; not recommended for NVE MD.
    use_device="cuda",
)
energy = atoms.get_potential_energy()
forces = atoms.get_forces()
stress = atoms.get_stress()
magmoms = atoms.get_magnetic_moments()

Set model to "0.3.0", "0.2.0", or "r2scan" to use pretrained CHGNet models.

Training

FastCHGNet supports structures, CIF files, JSON datasets, and precomputed graphs. Use ExtXYZData for extended XYZ files and ASELMDBData for ASE LMDB files. To train on an extended XYZ dataset, edit config/fastchgnet.yml and run:

torchrun --nproc_per_node=4 examples/train.py --config config/fastchgnet.yml

Use --config to specify the YAML file containing all training settings. For legacy MPtrj LMDB files storing raw VASP stress, set data_config.stress_unit: vasp_kbar.

Acknowledgements and license

FastCHGNet is based on the CHGNet source code by Deng et al. (Ceder Group). The original copyright notices and Modified BSD LICENSE (BSD-3-Clause-LBNL) are retained. FastCHGNet modifications are Copyright (c) 2026 Yuanchang Zhou under the same license.

Please also cite CHGNet. For r2scan, follow its model citation.

Reference

If you use FastCHGNet in your research, please cite:

@INPROCEEDINGS{zhou2025fastchgnet,
  author={Zhou, Yuanchang and Hu, Siyu and Wang, Chen and Wang, Lin-Wang and Tan, Guangming and Jia, Weile},
  booktitle={2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS)}, 
  title={FastCHGNet: Training One Universal Interatomic Potential to 1.5 Hours with 32 GPUs}, 
  year={2025},
  doi={10.1109/IPDPS64566.2025.00029}
}

@article{deng2023chgnet,
  title = {CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling},
  author = {Deng, Bowen and Zhong, Peichen and Jun, KyuJung and Riebesell, Janosh and Han, Kevin and Bartel, Christopher J. and Ceder, Gerbrand},
  journal = {Nature Machine Intelligence},
  year = {2023},
  doi = {10.1038/s42256-023-00716-3}
}