DOCSEARCHbyAlgolia
Submit Docs

fairchem

repository·main·Indexed 24 days ago

https://github.com/facebookresearch/fairchem

A centralized repository for FAIR Chemistry's data, models, and applications in materials science and quantum chemistry. It provides pretrained machine learning interatomic potentials (MLIPs), such as the UMA models, for tasks including catalysis, molecular dynamics, and inorganic materials relaxation. The library includes the FAIRChemCalculator for ASE integration and a CLI for running UMA benchmarks and evaluations across materials, OMC, catalysis, and NVE MD domains.

Tokens
100.4K
Snippets
211
Records
460
Agent score
81%

What's inside fairchem

  1. Overview of Open Catalyst Project datasets

    main

    The Open Catalyst Project provides several specialized datasets used for training and evaluating machine learning models in heterogeneous catalysis applications.

    Recommended Starting Point: For most new users, it is recommended to start with the UMA model, which has been pre-trained on all available catalyst datasets and offers state-of-the-art performance.

  2. Overview of Open Catalyst Experiments 2024 (OCx24)

    main
    Open Catalyst Experiments 2024 (OCx24) is a project designed to bridge the gap between computational descriptors and experimental outcomes in heterogeneous catalysis. It focuses on two key green chemistries: the hydrogen evolution reaction (HER) and the electrochemical reduction of carbon dioxide (CO2RR). The project provides curated datasets of experimental results obtained under industrially relevant conditions and models that predict these experimental outcomes using computational features.

  3. Overview of Universal Models & Architectures in FAIR Chemistry

    main

    The FAIR Chemistry team has developed several state-of-the-art neural network architectures for atomic property prediction. Key models include:

    • UMA (Universal Models for Atoms): A family of models trained on 500M+ structures with 1.4B parameters, utilizing a mixture of linear experts.
    • eSEN: Designed for learning smooth and expressive interatomic potentials, passing energy conservation tests and achieving SOTA on thermal conductivity and phonons.
    • EquiformerV2: An improved equivariant transformer that provides significant improvements in force and energy predictions.
    • eSCN: A model that reduces the complexity of SO(3) equivariant convolutions from $O(L^6)$ to $O(L^3)$.

  4. Overview of FAIR Chemistry Generative Models

    main

    The FAIR chemistry team provides several generative models designed to create novel materials and molecules by learning patterns from training data. These models enable rapid exploration of chemical space. Note that most of these models currently reside in separate repositories rather than the main fairchem repository.

    Available models include:

    • All-atom Diffusion Transformers (ADiT): A transformer-based latent diffusion approach that encodes both molecules and materials in the same latent space for synergistic learning.
    • FlowLLM: A hybrid approach using an LLM to generate chemical compositions and Flow Matching to generate the corresponding crystal structures.
    • FlowMM: Uses Flow Matching to efficiently generate inorganic crystal structures, offering faster generation than some prior diffusion-based methods.
    • Crystal-text-llm: Uses fine-tuned LLaMA models to generate molecules or materials as text, enabling text conditioning.

  5. Overview of the OMol25 dataset

    main

    The Open Molecules 2025 (OMol25) dataset is a large-scale collection of over 100 million single-point DFT calculations. It covers a broad domain including organic/inorganic molecules, transition metal complexes, and electrolytes.

    Key Specifications:

    • Labels: Total energy (eV) and forces (eV/A).
    • Level of Theory: wB97M-V/def2-TZVPD (calculated via ORCA6).
    • License: CC BY 4.0.
    • Download: Available via HuggingFace.

  6. Overview of the ODAC25 dataset

    main

    The Open DAC 2025 (ODAC25) dataset is a large-scale collection of approximately 70 million DFT single-point calculations designed for training machine-learned interatomic potentials for direct air capture applications. It focuses on the adsorption of CO2, H2O, N2, and O2 within nearly 15,000 metal-organic framework (MOF) structures.

    Key Specifications:

    • Domain: Metal-organic frameworks (MOFs)
    • Adsorbates: CO2, H2O, N2, O2
    • Labels: Total energy (eV) and forces (eV/Å)
    • Level of Theory: PBE+D3 (VASP)
    • License: CC BY 4.0
    • Download Source: HuggingFace

  7. Overview of ODAC classical force field files

    main

    The src/fairchem/data/odac/force_field/ directory contains the following resources for classical force field (FF) analysis:

    • data_w_oms.json: Contains successful FF interaction energy calculations with system information and DFT-computed interaction energies.
    • data_w_ml.json: Contains the same information as data_w_oms.json but specifically for systems with successful ML interaction energy predictions (limited to the in-domain test set).
    • FF_analysis.py: A script used to perform error calculations and generate analysis plots (specifically the four panels of Figure 5 from the associated paper).
    • FF_calcs/: A folder containing example calculations for classical FF interaction energy predictions.
    • in.lammps: A template LAMMPS input file used for all calculations.

  9. Overview of the Open Catalyst 2020 Dense (OC20Dense) dataset

    main
    The OC20Dense dataset is a validation dataset designed for assessing model performance in calculating global minimum adsorption energies. It contains 85,658 unique configurations sampled from approximately 1,000 adsorbate+surface materials. Unlike the standard OC20 dataset which focuses on local relaxations, OC20Dense is optimized for determining the global minimum energy site, which is critical for computational catalysis research under low coverage conditions.

  10. Overview of the Open Catalyst 2020 Multi-Adsorbate (mAds) Dataset

    main

    The OC20-mAds dataset is an expansion of the original OC20 dataset designed to study multi-adsorbate and coverage effects on catalyst surfaces. It contains approximately 21.8 million structures. In this dataset, adsorbates are randomly sampled from the OC20 adsorbate list, with a maximum of 5 adsorbates per surface. In some cases, all adsorbates on a surface may be identical.

    Key Specifications:

    • Size: 21.8M structures
    • Max Adsorbates: Up to 5 per surface
    • License: CC-BY-4.0
    • Reference: UMA Paper

  11. Overview of the OMat24 dataset

    main

    The Open Materials 2024 (OMat24) dataset is a collection of inorganic bulk materials structures. It contains approximately 1.07M training structures and 1.02M validation structures. The dataset includes single-point calculations of non-equilibrium structures and structural relaxations, labeled with:

    • Total energy (eV)
    • Forces (eV/A)
    • Stress (eV/A^3)

    Data is provided in ASE DB compatible LMDB files and is Matbench-Discovery compatible. The level of theory used is DFT (PBE/PBE+U).

  12. Overview of the OMC25 dataset

    main

    The Open Molecular Crystals 2025 (OMC25) dataset contains over 25 million structures of organic molecular crystals. These structures were generated from relaxation trajectories of random packings of OE62 molecules into 3D unit cells using the Genarris 3.0 package.

    Key Specifications:

    • Domain: Organic molecular crystals
    • Labels: Total energy (eV), forces (eV/A), and stress (eV/A^3)
    • Level of Theory: PBE+D3 (calculated via VASP)
    • License: CC BY 4.0
    • Format: ASE DB compatible .aselmdb files