graphite

A repository for implementing graph network models based on atomic structures.

Implemented or replicated works

• ALIGNN
  • Atomistic Line Graph Neural Network for improved materials property predictions
• ALIGNN-d (our work, see demo)
  • Efficient and interpretable graph network representation for angle-dependent properties applied to optical spectroscopy
• Gated GCN
  • Benchmarking Graph Neural Networks
• NequIP (implementation adopted from e3nn; requires e3nn installation)
  • E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials
• Atomic Structure denoiser (our work, see demo)
  • Score-based denoising for atomic structure identification
• MeshGraphNets
  • Learning Mesh-Based Simulation with Graph Networks
• Score dynamics (our work, see demo)
  • Score dynamics: scaling molecular dynamics with picoseconds time steps via conditional diffusion model
• Spectroscopy-guided generation of amorphous structures (our work, see demo)
  • Spectroscopy-guided discovery of three-dimensional structures of disordered materials with diffusion models
• Equivariant transformer
  • Equivariant pretrained transformer for unified geometric learning on multi-domain 3D molecules
• Graphormer
  • Towards predicting equilibrium distributions for molecular systems with deep learning

Gallery

Spectroscopy-guided amorphous material generation

Unconditional generation of amorphous carbons via diffusion model. Color is meant to help give you a sense of depth.

Conditional generation of amorphous carbons based on a given XANES spectrum.

Generation of multi-element, amorphous carbon nitrides.

Score dynamics

Generative rollouts of molecular dynamics over picosecond timesteps via conditional diffusion model.

Atomic structure denoiser

Simple and effective atomic denoiser for structure characterization.

Installation

The installation time is typically less than 10 minutes on a normal local machine.

Installation dependencies:

• pytorch>=2.0.1
• torch_geometric
• torch-scatter
• torch-cluster

Reasons: many model implementations in this repo are written mostly based on PyTorch, but some operations such as scatter and graph pooling are from PyTorch Geometric (PyG), which offers very optimized CUDA implementations. Additionally, most models in this repo treat atomic/molecular data as graphs following the PyG data format (torch_geometric.data.Data).

Also, this repo has scattering (via Torch.scatter_reduce) and clustering (e.g., radius_graph and knn_graph) codes such that torch-scatter and torch-cluster are not strictly required. Still, it is recommended to install torch-scatter and torch-cluster if you favor CUDA-optimized compute speed.

Lastly, for development purposes, you may want to install packages such as ase, MDAnalysis, rdkit, lightning, etc.

Example installation process:

conda create -n graphite
conda activate graphite
conda install pytorch pytorch-cuda=12.1 -c pytorch -c nvidia
pip install torch_geometric
pip install torch_scatter torch_cluster -f https://data.pyg.org/whl/torch-2.3.0+cu121.html

# For development depending on your use case
pip install ase jupyterlab ipywidgets seaborn lightning tensorboard

Then, to install graphite, clone this repo and run:

pip install -e /path/to/the/repo

To uninstall:

pip uninstall graphite

How to use

graphite is intended for research and prototyping. It is a general collection of simple codes containing helper functions, custom graph convolutions, model templates, and so on. Full-scale production codes for specific applications and deployments should be hosted elsewhere.

• The src folder contains the source code.
• The notebooks folder contains Jupyter notebooks that demonstrate running or training models.
  • Some demos require additional packages (e.g., PyTorch Lightning for automated training).
  • The demo notebooks are not always up-to-date. We will try to update the notebooks after every major change to the source code.
• The media folder contains media files.

Release

LLNL-CODE-836648