Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 2,138 words
Add missing citations, update stale details, or suggest a clearer explanation.
UMA (Universal Model for Atoms) is a family of machine-learning interatomic potentials released in 2025 by the FAIR Chemistry team at Meta AI. A UMA model predicts the potential energy, per-atom forces, and cell stresses of an arbitrary collection of atoms, and works as a single general-purpose model across chemistry, catalysis, and materials science rather than being tuned to one narrow problem. The intended role is to replace expensive quantum-mechanical calculations, in particular density functional theory, with a model that is roughly comparable in accuracy on many tasks while running orders of magnitude faster. The paper that describes the family, titled "UMA: A Family of Universal Models for Atoms," was posted to arXiv on June 30, 2025 (with a revised version on March 4, 2026), led by Brandon M. Wood and Muhammed Shuaibi. UMA was announced alongside the Open Molecules 2025 (OMol25) dataset in a Meta AI research post on May 14, 2025, and both the model weights and the supporting data were released openly on Hugging Face.
The core problem UMA addresses is the cost of atomistic simulation. Predicting how atoms move and bond normally requires solving an approximation of the quantum many-body problem, and DFT, the most common practical method, scales steeply with system size (roughly cubically with the number of electrons). A single small calculation can take minutes to hours, making long simulations or large screening campaigns prohibitively slow. A machine-learning interatomic potential learns a fast surrogate of that energy surface from a training set of DFT calculations, then evaluates new structures in a fraction of a second with cost that grows close to linearly with the number of atoms.
Earlier interatomic potentials were usually trained on small, problem-specific datasets, which limited how broadly any one model could be trusted. UMA was built around the opposite bet: that the data-and-compute scaling seen in language and vision models could carry over to atomistic chemistry. Meta reported training UMA on roughly 500 million unique three-dimensional atomic structures, which the authors describe as the largest such training effort to date, drawn from molecules, materials, catalyst surfaces, molecular crystals, and metal-organic frameworks. The result is positioned as a foundation model for atoms and as part of the broader AI for science effort, with the explicit goal of showing that one model can match or beat specialized models across many domains without per-task fine-tuning.
A machine-learning interatomic potential takes the elements and three-dimensional coordinates of a set of atoms and returns a single scalar energy. Because forces are the negative gradient of energy with respect to atomic positions, and stress is the gradient with respect to the simulation cell, a model that produces a smooth, differentiable energy can yield forces and stresses by automatic differentiation. Predicting forces this way (a "conservative" force field) is important for molecular dynamics, because energy-conserving forces keep simulations stable, whereas models that predict forces directly can drift over long trajectories.
Modern potentials, including UMA, are built as equivariant graph neural networks. Atoms are nodes, nearby atoms are connected by edges, and the network is constructed so that rotating or translating the input rotates the predicted forces and leaves the energy unchanged. This physical symmetry, known as equivariance, sharply reduces the amount of data needed to learn a reliable model. UMA builds on an architecture Meta published in February 2025 called eSEN (the equivariant Smooth Energy Network, arXiv 2502.12147), which emphasized smoothness so the learned energy surface is well-behaved enough to conserve energy during dynamics. eSEN itself reached state-of-the-art results on materials stability, thermal conductivity, and phonon benchmarks, and UMA inherits several of its components, including the eSCN convolution and smoothing envelope functions.
UMA is an equivariant graph neural network that updates spherical-harmonic node embeddings, extending the eSEN backbone. Its central new idea is a Mixture of Linear Experts (MoLE), which adapts the mixture-of-experts approach familiar from large language models to the interatomic-potential setting. Instead of routing individual tokens to different sub-networks, MoLE mixes a set of linear weight matrices using coefficients that depend on global, system-level properties: the total charge, the spin multiplicity, and which DFT task (that is, which dataset and level of theory) the input belongs to. Those global features are embedded and passed through a small multilayer perceptron that produces the mixing weights.
The key efficiency trick is that, because the experts are linear and the mixing weights are global to the structure, the chosen experts can be pre-merged into a single effective weight matrix before inference. This lets the model carry a large total parameter count for capacity while keeping the number of parameters actually used per structure small, so inference stays fast and equivariance is preserved inside the eSCN convolution. Meta reported that increasing from one expert to eight gave a large drop in loss, with diminishing returns by 32 and little benefit by 128 experts, and that MoLE outperformed naive multi-task training and even beat several single-task models, indicating useful transfer of knowledge across very different chemistries.
The family ships in three sizes. UMA-small (UMA-S) has roughly 150 million total parameters but only about 6 million active per structure (Meta also cites a 145M-total, 6M-active configuration); UMA-medium (UMA-M) has about 1.4 billion total parameters with roughly 50 million active; and UMA-large (UMA-L) has about 700 million parameters, all active (the small and medium variants are the ones that use MoLE). Training followed a two-stage recipe taken from eSEN: a first stage that predicts forces directly for faster training, then a fine-tuning stage that produces energy-conserving forces and stresses. Meta also fit empirical scaling laws relating compute, data, and model size to choose the configurations. On reported benchmarks, UMA-M reached an F1 of 0.930 on Matbench Discovery for materials stability, and UMA improved successful adsorption-energy calculations on the AdsorbML catalysis benchmark, in each case without task-specific fine-tuning. For speed, Meta reported UMA-S simulating 1,000 atoms at about 16 steps per second and fitting systems of up to 100,000 atoms in memory on a single 80GB GPU. UMA is distributed through Meta's open fairchem codebase under the FAIR Chemistry License, which permits academic and commercial use but excludes certain jurisdictions (China, Russia, and Belarus) and is governed by an acceptable-use policy.
UMA's molecular accuracy depends heavily on Open Molecules 2025 (OMol25), a high-accuracy quantum-chemistry dataset released by the same team (arXiv 2505.08762, posted May 13, 2025, led by Daniel S. Levine). OMol25 contains more than 100 million DFT single-point calculations, covering roughly 83 million unique molecular systems, all computed at the omegaB97M-V level of theory with the def2-TZVPD basis set using the ORCA program (version 6.0.1). Meta reported that generating it took on the order of 6 billion compute core-hours, making it one of the largest high-quality molecular quantum-chemistry datasets assembled. It spans 83 elements, includes explicit treatment of variable charge and spin, and extends to systems of up to 350 atoms, considerably larger than most prior molecular DFT datasets. Its three priority areas are biomolecules (drawn from sources such as the Protein Data Bank, including protein-ligand interfaces and unusual nucleic-acid structures), electrolytes (aqueous and organic solutions, ionic liquids, molten salts, and battery-relevant species), and metal complexes (combinatorially generated across metals, ligands, and spin states). It also recomputes several existing community datasets, including SPICE, Transition-1x, and ANI-2x, at the same level of theory. OMol25 is released under a CC-BY-4.0 license.
OMol25 is the newest entry in a lineage that began with the Open Catalyst Project. UMA was trained on five FAIR Chemistry datasets together, each carrying its own DFT task: OC20 (catalyst surfaces, 2020), OMat24 (inorganic materials, 2024), ODAC (metal-organic frameworks for direct air capture, originally Open DAC 2023), OMC25 (molecular crystals), and OMol25 (molecules). In combination these contain close to 500 million training examples and more than 30 billion atoms. In 2024 Meta consolidated these formerly separate efforts, originally branded under the Open Catalyst Project, into a single FAIR Chemistry codebase and brand named fairchem, which now hosts the datasets, the eSEN models, and UMA.
Because UMA produces energies, forces, and stresses for almost any atomic system, it can drive the same workflows that previously relied on DFT, but at far lower cost. Reported and intended applications include heterogeneous catalysis (screening adsorbates on metal and oxide surfaces, the original Open Catalyst use case), battery and electrolyte research (stability and degradation of electrolyte species), metal-organic frameworks and gas adsorption for carbon capture, inorganic materials discovery and stability prediction, molecular crystal structure prediction, and the study of drug-like and biomolecular systems relevant to early-stage drug discovery. Independent users have reported, for example, that an organometallic reaction-barrier study that once needed weeks of CPU time could be reproduced in an hour or two. The reported gains come with caveats that the authors themselves flag: accuracy degrades on strongly charged or open-shell (unpaired-electron) systems, where errors on properties like ionization energies can be large, and the models as released do not natively account for solvent effects, so some real-world chemistry still needs additional corrections.
UMA is one of the clearest demonstrations to date that the scaling-driven, single-foundation-model approach from deep learning and generative AI can transfer to atomistic simulation. By training one model on roughly half a billion structures spanning chemistry, materials, and catalysis, and by introducing the Mixture of Linear Experts to reconcile data computed at different levels of theory without slowing inference, Meta argued that a universal atoms model can match or beat narrow specialists across many benchmarks at once. Combined with the open release of both the weights and the very large OMol25 and related datasets, UMA lowered the barrier to fast, broadly applicable atomistic simulation and has been positioned by Meta as core infrastructure for AI-driven scientific discovery in chemistry and materials.
| Item | Detail |
|---|---|
| Developer | FAIR Chemistry team, Meta AI |
| Model type | Machine-learning interatomic potential (equivariant graph neural network) |
| Outputs | Energy, per-atom forces, cell stress |
| UMA paper | "UMA: A Family of Universal Models for Atoms," arXiv 2506.23971 (posted June 30, 2025; v2 March 4, 2026) |
| Public announcement | Meta AI blog, May 14, 2025 |
| Backbone | eSEN (equivariant Smooth Energy Network), arXiv 2502.12147 (Feb 17, 2025) |
| Key innovation | Mixture of Linear Experts (MoLE) conditioned on charge, spin, and DFT task |
| UMA-S | ~150M total parameters, ~6M active (also cited as 145M total) |
| UMA-M | ~1.4B total parameters, ~50M active |
| UMA-L | ~700M total parameters, all active |
| Training data | ~500 million atomic structures, >30 billion atoms |
| Source datasets | OC20, OMat24, ODAC (Open DAC), OMC25, OMol25 |
| Reported inference (UMA-S) | ~1,000 atoms at ~16 steps/sec; up to ~100,000 atoms on one 80GB GPU |
| Reported benchmark | Matbench Discovery F1 = 0.930 (UMA-M); state-of-the-art AdsorbML |
| Model license | FAIR Chemistry License (commercial use allowed; excludes some jurisdictions) |
| OMol25 paper | "The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models," arXiv 2505.08762 (May 13, 2025) |
| OMol25 size | >100M DFT calculations, ~83M unique systems, up to 350 atoms |
| OMol25 level of theory | omegaB97M-V / def2-TZVPD, computed with ORCA 6.0.1 |
| OMol25 scope | 83 elements; biomolecules, electrolytes, metal complexes; ~6 billion compute hours |
| OMol25 license | CC-BY-4.0 |
| Code / distribution | fairchem (formerly Open Catalyst Project), Hugging Face |