graph-pes

Date: Apr 22, 2025 - Author: John Gardner - Version: 0.1.0

graph-pes is a package designed to accelerate the development of machine-learned potential energy surfaces (ML-PESs) that act on graph representations of atomic structures.

The core component of graph-pes is the GraphPESModel. You can take any model that inherits from this class and:

• train and/or fine-tune it on your own data using the graph-pes-train command line tool

• use it to drive MD simulations via LAMMPS or ASE

We provide many GraphPESModels, including:

• re-implementations of popular architectures, including NequIP, PaiNN, MACE and TensorNet

• wrappers for other popular ML-PES frameworks, including mace-torch, mattersim, and orb-models, that convert their models into graph-pes compatible GraphPESModel instances

Use graph-pes to train models from scratch, experiment with new architectures, write architecture-agnostic validation pipelines, and try out different foundation models with minimal code changes.

Useful links:

🔥 Train

Train an existing architecture from scratch

Train a model

🔍 Analyse

Analyse a trained model

https://jla-gardner.github.io/graph-pes/quickstart/quickstart.html#Model-analysis

🔧 Fine-tune

Fine-tune a foundation model on your data

Fine-tuning

🔨 Build

Implement your own ML-PES architecture

Implement a model

🧪 Experiment

Define a custom training loop

Custom training loops

🎓 Learn

Learn more about the properties of PESs

Theory

Installation:

Install graph-pes using pip. We recommend doing this in a new environment (e.g. using conda):

conda create -n graph-pes python=3.10 -y
conda activate graph-pes
pip install graph-pes

Please see the GitHub repository for the source code and to report issues.