facebookresearch/fairchem

What it solves

Fairchem provides a centralized hub for data, models, and tools designed for materials science and quantum chemistry. It enables researchers to perform complex chemical simulations—such as relaxing adsorbates on catalytic surfaces, relaxing inorganic crystals, and running molecular dynamics—without needing to rely solely on computationally expensive traditional methods.

How it works

The project provides pretrained machine learning interatomic potentials (MLIPs), specifically the UMA model series (small and medium versions), which can be integrated into the Atomic Simulation Environment (ASE) via the FAIRChemCalculator. Users select a specific task name (e.g., oc20 for catalysis, omat for inorganic materials, omol for molecules and polymers) to tailor the model's predictions to their specific domain. For large-scale simulations, it supports multi-node and multi-GPU inference and integrates with LAMMPs to handle systems with hundreds of thousands of atoms.

Who it’s for

It is designed for scientists and researchers in materials science, quantum chemistry, and computational chemistry who need to predict potential energy and perform structural optimizations or molecular dynamics simulations.

Highlights

• Broad Domain Coverage: Supports catalysis, oxide catalysis, inorganic materials, molecules, polymers, MOFs, and molecular crystals.
• UMA Model Series: Offers a range of models from the fast uma-s (SOTA on most benchmarks) to the more accurate uma-m.
• High-Performance Scaling: Supports multi-GPU and multi-node inference via Ray, enabling simulations of systems with 100k+ atoms.
• ASE Integration: Seamlessly integrates with the ASE library for easy setup of atoms objects and dynamics objects.