Open Catalyst Project

Overview

The repository documents fairchem, the FAIR Chemistry team's collection of models, data and applications for materials science and quantum chemistry. For this Open Catalyst Project entry, the relevant documented functionality is its catalysis workflow: pretrained UMA models are accessed through fairchem-core and connected to the Atomic Simulation Environment (ASE) through FAIRChemCalculator. The excerpt describes the current fairchem workflow rather than providing a separate account of Open Catalyst Project datasets or legacy models.

Inputs are atomic structures represented as ASE atoms objects, paired with a model and a domain-specific task name. The README assigns oc20 to catalysis, oc22 to oxide catalysis and oc25 to (electro)catalysis, with model-family restrictions noted for the latter two. Its surface example builds a copper slab with an adsorbed CO molecule and uses an ASE optimizer to relax the structure. Other examples demonstrate inorganic crystal relaxation, molecular dynamics with trajectory output, and molecular energy differences between spin states. These place the predictive models inside simulation workflows rather than presenting fairchem as a standalone discovery agent.

Practical constraints include applying for UMA model access through Hugging Face and authenticating before using the pretrained models. Multi-GPU and multi-node inference, including LAMMPS integration, require Ray. For inorganic comparisons, the README warns that OMat24-trained model energies are not directly compatible with Materials Project calculations and require attention to reference energies and corrections. Repository code is MIT-licensed, but checkpoint licences vary by application area; the code licence should not be treated as permission for every model or dataset.

Key Features

• ASE integration through FAIRChemCalculator connects pretrained UMA predictions to structure optimization and molecular dynamics workflows.
• Domain-specific task selection includes oc20 for catalysis, oc22 for oxide catalysis and oc25 for (electro)catalysis; the README marks oc22 and oc25 as restricted to 1p2 models.
• A documented adsorbate-relaxation example combines a copper surface, CO adsorbate and ASE's LBFGS optimizer.
• Additional examples cover inorganic crystal relaxation, molecular dynamics trajectory writing and molecular spin-gap calculation using charge and spin metadata.
• Multi-GPU and multi-node inference uses a workers setting and supports LAMMPS workflows, with Ray required.

Use Cases

• Suggested evaluation: adapt the documented CO-on-copper relaxation example to representative adsorbate–surface systems and compare the relaxed structures with suitable reference calculations.
• Suggested evaluation: assess the appropriate catalysis task for oxide or electrochemical surface structures, checking the documented model restrictions before inference.
• Suggested evaluation: explore ASE-driven relaxation or dynamics as a preliminary stage in a materials-discovery workflow, validating relevant energy differences against consistently configured references.

How to Use

1. Read the fairchem README to identify the pretrained-model workflow relevant to your system. The supplied source describes fairchem and UMA rather than a separate Open Catalyst Project quick start.
2. Follow the README's fairchem-core installation instructions in an isolated environment. Choose the standard workflow or consult the additional requirements for distributed inference; do not assume every optional integration is installed.
3. Apply for access at the UMA model repository and complete the documented Hugging Face authentication process. Check the applicable checkpoint terms separately from the repository code licence.
4. Prepare an ASE atoms object and select the appropriate task. For catalysis, start with the documented oc20 surface-relaxation example; verify model restrictions before selecting oc22 or oc25.
5. Attach FAIRChemCalculator and adapt the documented optimizer or dynamics example. Evaluate outputs against suitable reference calculations. For distributed inference or LAMMPS, consult the common-task documentation and its Ray requirements.