FAIR-Chem / OpenCatalyst Gardens

OpenCatalyst Garden is a repository of state-of-the-art machine learning models developed by the FAIR-Chem team for catalyst discovery. These models are trained on the Open Catalyst Project (OC20) dataset and published on Garden, making them easily accessible and reproducible.

Overview

The Open Catalyst Project (OC20) provides a large-scale dataset of DFT calculations for catalyst surface reactions. This repository contains Modal-based deployments of several models, enabling rapid structure optimization and energy/force predictions, which are essential for catalysis research.

Quick Start

Follow these simple steps to get started locally:

1. Clone the repository:

   git clone https://github.com/Garden-AI/opencatalyst-garden.git
   cd opencatalyst-garden

2. Install dependencies using uv:
   (Ensure you have Python 3.10 and that uv is installed.)

   uv pip install .

3. Run the Modal app:

   uv run modal run src/OC20_modal_app.py

Available Models

• EquiformerV2
• GemNet-OC
• eSCN
• PaiNN
• SchNet
• DimeNet++
• SCN

Available Tasks

1. Structure to Energy and Forces (S2EF)

   • Predicts atomic forces and total energy for a given structure.
   • Useful for understanding local atomic interactions.

2. Initial Structure to Relaxed Energy (IS2RE)

   • Predicts the relaxed energy of a structure without performing full relaxation.
   • Efficient for screening many candidate structures.

Model Architectures

• EquiformerV2: Transformer-based model with E(3)-equivariant layers.
• GemNet-OC: Graph neural network optimized for OC20.
• eSCN/SCN: Spherical Channel Networks.
• PaiNN: Polarizable interaction neural network.
• SchNet: Continuous-filter convolutional neural network.
• DimeNet++: Directional message passing neural network.

Modal App Definition

Check out the Modal App definition in OC20_modal_app.py to see how the models are loaded and configured to run on Modal's GPUs.

Demo Notebook

For a detailed example of how to use these models for catalyst structure prediction and analysis, check out the FairChem_OCP_Garden_Demo.ipynb notebook.

Contributing

Contributions are welcome! If you have suggestions or spot issues, please feel free to submit a Pull Request or open an issue.

Acknowledgments

• The Open Catalyst Project team for the dataset and baseline models.
• The FAIR-Chem team for model development and training.