FAIR Chemistry and UMA: A Cross-Domain Atomic Foundation Model Platform
New1. Overview
The facebookresearch/fairchem repository is the consolidated codebase of the Meta FAIR Chemistry team. It was created on September 26, 2019, last pushed on November 12, 2026, uses main as its default branch, is written primarily in Python, and publishes documentation at facebookresearch.github.io/fairchem. As of November 2026 the repository carries roughly 2356 stars, 529 forks, and 38 open issues.
The name invites a misconception. In its early phase the project centered on the OC20 electrocatalysis dataset, training graph neural networks such as GemNet-OC, eSCN, and EquiformerV2 so that machine learning potentials could replace first-principles relaxations when computing adsorption energies. The repository today has grown into a broad platform for materials science and quantum chemistry, with data, models, demo code, and applications all living in one tree that is split into multiple publishable subpackages.
The flagship model is UMA. Its preprint, titled "UMA: A Family of Universal Models for Atoms", was posted in June 2025 as arXiv 2506.23971 with eighteen authors and Brandon M. Wood listed first. UMA is an equivariant graph neural network with a mixture-of-linear-experts architecture, trained on five hundred million unique three-dimensional structures assembled across the molecular, materials, and catalyst domains. A single model serves seven different levels of theory simply by switching the task name.
For practitioners this shift matters. The selection question is no longer which OC20 model to use, but which domain the system falls into, which DFT level of theory applies, and how license and weight access constraints interact. The remaining sections work through that chain.
2. How It Works
The backbone of UMA is an equivariant graph neural network built around the mixture of linear experts, abbreviated MoLE. It uses a single output head with dynamic routing, which allows the total parameter count to grow without slowing inference. The paper reports that UMA-medium holds one point four billion parameters while only about fifty million are active for any given atomic structure. That trade-off lets one model cover a wide chemical space while remaining affordable inside the repeated evaluation loop of a molecular dynamics run.
Routing requires four kinds of input. The task name determines which level of theory is emulated. The total charge of the system applies only to the omol task and defaults to zero. The total spin multiplicity applies only to omol as well and defaults to one. The unordered elemental composition is represented by an atom embedding per element, with the composition embedding taken as the mean over all atom embeddings.
The mapping from task to theory level is the core of the system. The omol task corresponds to the omegaB97M-V/def2-TZVPD level from OMol25, implemented in ORCA 6 with non-local dispersion and explicit solvation. The omc task corresponds to the PBE+D3 level from OMC25, implemented in VASP and aimed at molecular crystals. The omat task corresponds to the PBE and PBE+U levels from OMat24, using Materials Project recommended settings and VASP 54 pseudopotentials without dispersion. The oc20 task corresponds to RPBE with VASP 5.4 pseudopotentials. The odac task corresponds to PBE+D3 from ODAC23. The oc25 task corresponds to RPBE+D3 with VASP 6.4 pseudopotentials and dipole corrections, and is available only from UMA 1.2. The oc22 task corresponds to spin-polarized PBE+U and is likewise available only from UMA 1.2. The README also notes that OC20 was later recomputed to produce total energies rather than adsorption energies.
Two inference modes exist. The high-precision mode keeps full FP32 with model merging and compilation. The nitro mode became the default in October 2026 and is deeply optimized for single-GPU molecular dynamics and relaxations. It runs on NVIDIA Ampere and newer architectures, with dedicated optimization for the Hopper generation and above in single-GPU single-system setups, and falls back to other modes with a warning on unsupported hardware. The former turbo preset has been removed and multi-worker workloads now use multigpu. The batch server defaults to batch. A separate DFTD3Calculator adds GPU-accelerated DFT-D3(BJ) dispersion energies, forces, and analytic stresses to any ASE calculator, with pbe and r2scan presets.
Training-side work matters as much. DeNS adds denoising of non-equilibrium structures as an auxiliary task. JMP trains on roughly 120 million systems from OC20, OC22, ANI-1x, and Transition-1x inside a multitask framework and reports a fifty-nine percent average improvement over training from scratch. FINETUNA is an online active learning framework that reports a ninety-one percent reduction in first-principles calculations while meeting accuracy thresholds ninety-three percent of the time. Conformal prediction provides calibrated uncertainty estimates. Together these explain how a single model can approach or exceed specialized models without per-task fine-tuning.
3. Scope and Assumptions
Several hard constraints should be confirmed before any work begins.
Weight access is gated. Users need a Hugging Face account, must request access to the facebook/UMA model repository, and must then authenticate with huggingface-cli login. In offline environments or corporate networks this sequence becomes a blocker.
Version compatibility has explicit edges. UMA 1.0 checkpoints are no longer supported and raise a RuntimeError on loading, because their mixture-of-experts include_self behavior diverges from later releases. Working with a 1.0 checkpoint requires installing fairchem-core at version 2.21.0 or earlier. Loading now also requires a defined model_id. UMA 1.1 ships without one and is tagged UMA-1.1 in memory, while UMA 1.2 and later already carry it; a checkpoint with neither is rejected. During training, model_id should be set on the model configuration.
Hardware has a firm floor. Nitro covers the Ampere, Hopper, and Blackwell generations only, and falls back to a non-nitro path on other cards. The project also notes that once a single step enters the five to ten millisecond range, the ASE integrator itself becomes the bottleneck, and a non-Python GPU integrator is recommended to reach the true performance of the models.
Task selection determines how much of the output can be trusted. All omol training data is aperiodic, so periodic systems require caution and inorganic materials tend to be poorly served. The omat task includes spin polarization effects but does not allow selecting a magnetic state, and OMat24 did not fully sample possible spin states. The oc20 and oc22 tasks contain neither oxides nor explicit solvents, and oc22 has no explicit solvent at all, so electrolyte environments need separate validation. The odac task covers only carbon dioxide and water adsorption, so hydrocarbon behavior inside MOFs falls outside the training range, and the number of bare MOF structures is limited. The oc25 model behaves surprisingly well on charged systems despite not receiving charge information, but work functions are not provided and require a subsequent first-principles calculation.
The two fine-tuned variants carry their own taboos. The r2scan variant is not spin polarized and describes every system as non-magnetic, so it should not be used for magnetic systems or cohesive energies, since isolated atoms are also often strongly magnetic; the small model with the omat task serves that purpose instead. The d4rev variant already contains dispersion and should not be combined with an additional D3 correction. Both fine-tunes train on less data than the original UMA-S-1.2.1, which is often more accurate at predicting DFT labels at its own level of theory. No isolated-atom reference energies ship with the models, so single-atom energies are unreliable.
4. Getting Started
Two installation routes exist. To use the published package, run pip install fairchem-core. To modify the source, clone the repository and install the development extras of fairchem/packages/fairchem-core in editable mode. The project recommends pairing this with a package manager and virtual environment such as uv.
The multi-package layout deserves attention. Under packages there is fairchem-core for core inference and training, fairchem-core-numpy126 for NumPy 2.6 compatible builds, fairchem-lammps for the LAMMPS interface, fairchem-demo-ocpapi for API demos, the data packages fairchem-data-oc, fairchem-data-odac, fairchem-data-omat, fairchem-data-omc, and fairchem-data-omol that generate configurations and VASP input sets, plus four application packages: AdsorbML, cattsunami, fastcsp, and ocx. The data packages are published separately on PyPI, with fairchem-core reaching 2.25.0 on October 27, 2026, fairchem-lammps at 0.4.0, fairchem-data-omol at 0.1.2, fairchem-data-odac and fairchem-data-omc at 0.1.0, and fairchem-data-omat at 0.2.0.
The minimal runnable path has three steps. Obtain a prediction unit with pretrained_mlip.get_predict_unit, wrap it in a FAIRChemCalculator with an explicit task_name, and hand an ASE Atoms object to the calculator. This interface covers the main workflows in the documentation: adsorbate relaxation on a catalytic surface, inorganic crystal relaxation, molecular dynamics, spin gap calculation, dispersion correction, and multi-GPU inference with LAMMPS.
Measured throughput is published. The horizontal axis of the benchmark figure is the daily nanoseconds implied by the mean runtime of an ASE NVE step on a 1000-atom silicon cell using an NVIDIA H200, at a one femtosecond timestep. The vertical axis is the force mean absolute error on the public OMol25 validation set in meV per angstrom. On a single H200 the nitro mode reaches about six nanoseconds per day for 1000 atoms in ASE molecular dynamics, raw inference throughput is equivalent to about thirty-three nanoseconds per day at 150 atoms, and systems up to fifteen thousand atoms are supported. The benchmark details and reproduction scripts live in the benchmarks/omol directory. Finer accuracy comparisons are available through the FAIR Chemistry leaderboard on Hugging Face.
5. Comparison with Alternatives
The positioning of UMA becomes clearer when set against neighboring lines of work.
The MACE line takes a materials-potential route. Its paper, titled "A foundation model for atomistic materials chemistry", appeared in the Journal of Chemical Physics, volume 163, on November 13, 2025, and emphasizes training a universal interatomic potential on Materials Project data that reaches first-principles accuracy after fine-tuning on a handful of application-specific points. Both routes pursue generalization; the difference lies in task coverage and the diversity of theory levels.
Specialized task models still carry value. GemNet-OC improved on the previous best on OC20 by sixteen percent while cutting training time to a tenth. eSCN reduced the complexity of equivariant convolution from the sixth to the third power of the representation degree. EquiformerV2 improved force error by up to nine percent and energy error by four percent on OC20, halved the number of first-principles calculations needed for adsorption energies, and matched or beat GemNet-OC trained on both OC20 and OC22 using OC22 alone. Specialized models retain more controllable precision on a single task, while the value of a general model is removing the per-system effort of data collection and training.
eSEN, arXiv 2502.12147, takes a different validation route. It attributes the gap between low test error and good physical property prediction to failures of energy conservation, proposes an energy conservation test first, and then improves model design accordingly, reporting results on materials stability, thermal conductivity, and phonons. The lesson for users is that test error alone does not establish fitness.
The generative direction plugs into the same ecosystem. ADiT generates periodic materials and non-periodic molecules in one unified latent diffusion framework. FlowMM extends Riemannian flow matching to crystals and FlowLLM combines language models with flow matching. FastCSP combines random structure generation with UMA-driven relaxation and free energy calculations.
The relationship to first-principles computation deserves precise wording. Machine learning potentials do not replace electronic structure methods; they take over the repetitive relaxation cycle, concentrating the first-principles budget on validating the leading candidates. This belongs to the same line of inquiry as surrogate modeling and physics foundation models.
6. Benchmarks and Evidence
The data ecosystem supporting UMA is large, and each dataset has its own scope that should be recorded separately.
OC20 contains 1,281,040 first-principles relaxations corresponding to roughly 264,890,000 single point evaluations across materials, surfaces, and adsorbates, and it is the foundational dataset of the Open Catalyst project. OC22 contains 62,331 relaxations and about 9,854,504 single point calculations spanning oxide materials, coverages, and adsorbates, built for oxygen evolution reaction catalyst development. OC25 holds 7,801,261 calculations spanning 1,511,270 unique explicit solvent environments, covering 88 elements with commonly used solvents and ions, and is the largest solid-liquid interface dataset available. OMat24 contains over 110 million first-principles calculations focused on structural and compositional diversity, with the accompanying EquiformerV2 models reaching a F1 score above 0.9 on Matbench Discovery. OMC25 contains over 27 million molecular crystal structures covering 12 elements with up to 300 atoms per unit cell, generated from dispersion-inclusive relaxation trajectories of more than 230,000 randomly generated structures built from roughly 50,000 organic molecules. ODAC23 contains more than 38 million calculations on more than 8,400 MOFs. ODAC25 contains nearly 60 million single point calculations for carbon dioxide, water, nitrogen, and oxygen adsorption in 15,000 MOFs. OPoly26 contains more than 6.57 million calculations on polymer clusters with over 1.2 billion total atoms. OCx24 obtained DFT-verified adsorption energies on about 20,000 materials through 685 million AI-accelerated relaxations, together with 572 synthesized samples and 441 gas diffusion electrodes for carbon dioxide reduction and hydrogen evolution.
Model-side figures include the training scale of five hundred million unique three-dimensional structures, described in the paper as the largest training run to date, and the UMA 1.2 release reported as roughly fifty percent faster and roughly forty percent more accurate on the Open Molecules test set, with expanded data coverage for catalysts, molecules, and polymers.
Application-level numbers are concrete. AdsorbML reports 87.36 percent accuracy in identifying low-energy adsorbate surface configurations together with about a two-thousand-fold speedup, and introduces the Open Catalyst Dense dataset with roughly 1,000 surfaces and 100,000 configurations. CatTSunami reports that a graph neural network potential trained on OC20 finds transition states within 0.1 eV of DFT ninety-one percent of the time with a speedup of twenty-eight times, and reproduces a reaction network of 61 intermediates and 174 reactions at DFT resolution in 12 GPU days against 52 GPU years. The Catlas framework explored 947 binary intermetallics for syngas-to-oxygenates conversion and identified 144 candidate materials including Pt-Ti, Pd-V, Ni-Nb, and Ti-Zn.
A cross-domain example appears in enzyme catalysis. A September 2026 paper applies the eSEN-omol potential to fully atomistic, quantum-accurate treatment of complete enzymes in explicit solvent with up to 54 thousand atoms and one microsecond of total simulation time, reproducing experimental barrier trends for the Claisen rearrangement in chorismate mutase, resolving critical intermediate states in PETase catalyzed polymer depolymerization, and distinguishing mechanistic alternatives for metal-activated phosphoryl transfer, with a reported thousand-fold speedup over typical hybrid calculations.
7. Role in Physics-AI Workflows
Three patterns dominate when UMA is placed inside a working process: substitution, closed loop, and embedding.
Substitution is the most direct. Where every candidate surface-adsorbate configuration previously required a full relaxation plus energy evaluation, the model now performs the relaxation and only shortlisted candidates receive first-principles validation. This decouples the compute budget from the combinatorial size of alloy and dopant searches.
Closed loop means active learning. FINETUNA exploits prior information from pretrained models to guide online sampling, reducing the number of required first-principles calculations, while conformal prediction supplies calibrated uncertainty so that sampling points are distributed more effectively. A model can also serve as the initial proposer inside an active learning cycle, and there a broad but lower-precision initial pool often costs less than starting from zero.
Embedding means treating the model as one component of a larger flow. In solid-liquid electrocatalysis, models trained on OC25 handle cells larger than 800 atoms with enhanced sampling up to seven nanoseconds to compute free-energy profiles for carbon monoxide dimerization on copper surfaces, varying surface charge, cation identity, and facet. That work finds dimerization weakly sensitive to charge and cation identity, with appreciable stabilization only at the most negative charge densities, and identifies a more favorable pathway on the stepped Cu(310) surface at modest reducing potentials. Such calculations are out of reach at ab initio scale.
Further concrete scenarios exist. For direct air capture, the ODAC series supports MOF adsorbent screening, and ODAC25 improves calculation accuracy and the treatment of flexible MOFs while providing adsorption energy and Henry's law coefficient predictions. For molecular crystal structure prediction, FastCSP combines random structure generation with UMA-driven relaxation and free energy calculations, so a single system can be resolved within hours on tens of GPUs. Crystal prediction and structural relaxation connect to the materials discovery chain and stand in a downstream relation rather than a competing one with convex-hull search efforts.
From a platform perspective, the repository is also an entry layer for attaching machine learning potentials to an existing numerical stack. Work on discretization and solver efficiency and work on propagating potential uncertainty into inverse problems can both draw on the conformal prediction capability of the models.
8. Limitations and Failure Modes
Several usage patterns return results that cannot be relied upon.
Task mismatch is the primary risk. Handing a periodic inorganic system to a molecular-data omol task, or handing an electrolyte environment to an oc20 or oc22 task that contains no explicit solvent, produces numbers without physical meaning, merely numerically computable. Magnetic systems and cohesive energies routed to a non-spin-polarized r2scan fine-tune deviate systematically.
Errors of the training level of theory are inherited. The model reproduces the behavior of the DFT level it was trained on, so the gap between PBE and r2scan, or between omegaB97M-V and RPBE, appears directly in the output. Treating PBE-level output as close to experimental truth leaves an error far larger than the model's own test error.
Some properties simply are not provided. Work functions, charge transfer, reaction mechanism paths, and the energy reference for absolute free energies all require a subsequent first-principles calculation. The models ship no isolated-atom reference energies, so single-atom energies and anything derived from them, such as atomization enthalpy, are unreliable.
Sampling and integrators are an engineering failure point. Once a single step enters the five to ten millisecond range, the bottleneck moves from the model to the ASE Python integrator, and the observed throughput falls noticeably below the model capability. Such problems are easily misdiagnosed as weak model performance.
Licensing and access are practical constraints. Weights are gated and require an application, and a commercial project must first separate gated model terms from the MIT code license, as well as the GnuV2 declaration of the fairchem-lammps package on PyPI from the MIT statement at the repository root. The datasets carry their own terms, so redistributing data alongside the model requires separate verification.
Engineering compatibility demands that checkpoint and library versions match. A UMA 1.0 checkpoint raises an error on a new library, and different tasks use different pseudopotential versions, with OC20 on VASP 5.4 and OC25 on VASP 6.4, so reusing calculation settings across tasks easily produces incomparable results.
9. Version and Source Notes
Several discrepancies should be recorded explicitly to avoid mixing them when citing.
Licensing has three layers. The LICENSE.md at the repository root is the standard MIT text with copyright attributed to Meta, Inc. and its affiliates, while the GitHub API license field returns NOASSERTION, which disagrees with the file and should not be preferred in citations. The fairchem-lammps package declares GnuV2 on PyPI, which differs from the top-level repository statement. The README license section states that the code is MIT while model and checkpoint licenses vary by application area. The UMA_NITRO_LICENSE_PROVENANCE.md file provides an itemized provenance for third-party components: the in-house UMA Nitro code is MIT, the embedded dense matrix multiplication examples come from NVIDIA CUTLASS v4.7.0 under BSD-3-Clause, NumPy and PyTorch are BSD-3-Clause, Triton is MIT, the CUTLASS and CuTe DSL runtime falls under the NVIDIA software license agreement, the CUDA Python bindings fall under the NVIDIA CUDA software license, and the UMA checkpoints are licensed under gated model terms independently of the FairChem MIT license.
Parameter counts also have two versions of the figure. The README model table gives uma-s-1p2p1 as 6.6 million active and 290 million total parameters, and uma-m-1p1 as 50 million active and 1.4 billion total. The documentation site describes six million active and 145 million total. The paper abstract gives 1.4 billion parameters for UMA-medium with roughly fifty million active per atomic structure. Citations should state which document is being used.
The relation between task and model carries version constraints. The oc22 and oc25 tasks are provided only from UMA 1.2 onward. The dataset documentation labels the OC25 card with the title of a solid-liquid interface dataset paper, while the current arXiv title at that identifier is the copper-surface carbon dioxide reduction study; the OC25 dataset is introduced within that same work, so the identifier points to one and the same publication.
The timeline can be recorded as follows. UMA and OMol first shipped in June 2025. Multi-node multi-GPU and LAMMPS interfaces arrived in October 2025. UMA 1.2 was released in March 2026. An interactive UMA playground demo followed in June 2026. A paper on quantum-accurate enzyme catalysis was published in September 2026. In October 2026 UMA Nitro became the default inference mode, alongside two UMA-S-1.2.1 fine-tuned models and GPU-enabled DFT-D3(BJ) dispersion. On the code side the latest PyPI release of fairchem-core is 2.25.0, published October 27, 2026.
For archival citation, the Zenodo record for the repository is 10.5281/zenodo.15587498, titled facebookresearch/fairchem: fairchem-core-2.2.0, with rights marked CC BY 4.0. That snapshot corresponds to version 2.2.0 and is not the same point in time as the current PyPI release.
10. References
- Wood B. M., Dzamba M., Fu X., et al. UMA: A Family of Universal Models for Atoms. arXiv:2506.23971, 2025. https://arxiv.org/abs/2506.23971
- Meta FAIR Chemistry. fairchem repository (code, weight access, data, and application packages). https://github.com/facebookresearch/fairchem
- Meta FAIR Chemistry. FAIR Chemistry papers overview (models, datasets, generative models, sampling, applications, training methods). https://github.com/facebookresearch/fairchem/blob/main/docs/core/fair_chemistry_papers.md
- Meta FAIR Chemistry. UMA model documentation (task to theory mapping, fine-tune limitations). https://facebookresearch.github.io/fairchem/uma
- Meta FAIR Chemistry. UMA release changelog (default inference mode migration, version compatibility, 1.2 features). https://github.com/facebookresearch/fairchem/blob/main/docs/core/uma_changelog.md
- Meta FAIR Chemistry. UMA Nitro license provenance note. https://github.com/facebookresearch/fairchem/blob/main/UMA_NITRO_LICENSE_PROVENANCE.md
- Sahoo S. J., Maraschin M., Varley J. B., et al. Insights into CO dimerization at electrified Cu interfaces from large-scale machine learning simulations (introducing the OC25 dataset). arXiv:2509.17862, 2025. https://arxiv.org/abs/2509.17862
- Levine D. S., Shuaibi M., Spotte-Smith E. W. C., et al. The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models. arXiv:2505.08762, 2025. https://arxiv.org/abs/2505.08762
- Barroso-Luque L., Shuaibi M., Fu X., et al. Open Materials 2024 (OMat24) Dataset. arXiv:2410.12771, 2024. https://arxiv.org/abs/2410.12771
- Chanussot L., Das A., Goyal S., et al. Open Catalyst 2020 (OC20) Dataset and Community Challenges. ACS Catalysis, 11(10), 6059-6072, 2021. https://doi.org/10.1021/acscatal.0c04525
- Tran R., Lan J., Shuaibi M., et al. The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts. ACS Catalysis, 13(5), 3066-3084, 2023. https://doi.org/10.1021/acscatal.2c05426
- Sriram A., Brabson L. M., Yu X., et al. The Open DAC 2025 (ODAC25) Dataset. arXiv:2508.03162, 2025. https://arxiv.org/abs/2508.03162
- Sriram A., Choi S., Yu X., et al. The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture. ACS Central Science, 10(5), 923-941, 2024. https://doi.org/10.1021/acscentsci.3c01629
- Gharakhanyan V., Barroso-Luque L., Yang Y., et al. Open Molecular Crystals 2025 (OMC25) Dataset. arXiv:2508.02651, 2025. https://arxiv.org/abs/2508.02651
- Liao Y.-L., Wood B., Das A., Smidt T. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. arXiv:2306.12059, 2023. https://arxiv.org/abs/2306.12059
- Passaro S., Zitnick C. L. eSCN: Reducing SO(3) Convolutions to SO(2). arXiv:2302.03655, 2023. https://arxiv.org/abs/2302.03655
- Gasteiger J., Shuaibi M., Sriram A., et al. GemNet-OC: Graph Neural Networks for Large Datasets. arXiv:2204.02782, 2022. https://arxiv.org/abs/2204.02782
- Fu X., Wood B. M., Barroso-Luque L., et al. eSEN: Learning Smooth and Expressive Interatomic Potentials. arXiv:2502.12147, 2025. https://arxiv.org/abs/2502.12147
- Malosso C., Bigi F., Pegolo P., et al. High-quality, high-information datasets for universal atomistic machine learning (MAD-1.6). arXiv:2603.02089, 2026. https://arxiv.org/abs/2603.02089
- Gao M., Shayesteh Zadeh A., Seal A., et al. Quantum-accurate atomistic modeling of enzyme catalysis using a machine learned potential. arXiv:2609.09293, 2026. https://arxiv.org/abs/2609.09293
- Batatia I., Benner P., Chiang Y., et al. A foundation model for atomistic materials chemistry. Journal of Chemical Physics, 163, 2025. https://doi.org/10.1063/5.0297006
- Meta FAIR Chemistry. fairchem-core archive snapshot 2.2.0. Zenodo, 2025. https://doi.org/10.5281/zenodo.15587498