graph-pes is a framework built to accelerate the development of machine-learned potential energy surface (PES) models that act on graph representations of atomic structures.

Links: Google Colab Quickstart - Documentation - PyPI

Features

• Experiment with new model architectures by inheriting from our GraphPESModel base class.
• Train your own or existing model architectures (e.g., SchNet, NequIP, PaiNN, MACE, TensorNet, etc.).
• Use and fine-tune foundation models via a unified interface: MACE-MP0, MACE-OFF, MatterSim, GO-MACE and Orb v2/3.
• Easily configure distributed training, learning rate scheduling, weights and biases logging, and other features using our graph-pes-train command line interface.
• Use our data-loading pipeline within your own training loop.
• Run molecular dynamics simulations with any GraphPESModel using torch-sim, LAMMPS or ASE

Quickstart

pip install -q graph-pes
wget https://tinyurl.com/graph-pes-minimal-config -O config.yaml
graph-pes-train config.yaml

Alternatively, for a 0-install quickstart experience, please see this Google Colab, which you can also find in our documentation.

Contributing

Contributions are welcome! If you find any issues or have suggestions for new features, please open an issue or submit a pull request on the GitHub repository.

Citing graph-pes

We kindly ask that you cite graph-pes in your work if it has been useful to you. A manuscript is currently in preparation - in the meantime, please cite the Zenodo DOI found in the CITATION.cff file.