Universal Atomistic Models: UMA Overview

• Universal Models for Atoms (UMA) are pretrained, graph-based ML interatomic potentials designed to deliver accurate zero-shot predictions across diverse atomistic systems.
• UMA employs a Mixture of Linear Experts (MoLE) mechanism and scaling laws to balance model capacity with simulation speed and parameter efficiency.
• UMA leverages massive cross-domain datasets and uncertainty metrics to ensure reliable performance in applications from catalysis to crystal structure prediction.

Searching arXiv for relevant UMA and precursor papers to ground the article in current literature. Universal Models for Atoms (UMA) denotes a family of universal, pretrained graph-based machine-learned interatomic potentials (MLIPs) developed as general-purpose atomistic foundation models for molecules, materials, catalysts, molecular crystals, and metal–organic frameworks. In the modern formulation, UMA is designed to test whether a single universal model can achieve strong zero-shot performance across these domains while remaining fast enough for practical simulation workloads such as molecular dynamics, relaxations, and structure search; the central claim is that a single model without any fine-tuning can perform similarly or better than specialized models (Wood et al., 30 Jun 2025). In a broader historical sense, the UMA agenda also includes earlier atom-centered and local-environment-centered approaches that treated transferable atomic environments, rather than whole molecules or crystals, as the basic learning unit, most notably the AML/amons framework (Huang et al., 2017).

1. Conceptual basis of universal atomistic modeling

A recurring idea across UMA-style work is that atomistic prediction can be organized around transferable local environments. In the AML/amons framework, the total energy is written as a sum of atomic contributions,

$E=\sum_I E^I,$

and the machine-learning model is posed at the atomic level,

$$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$

In the Gaussian-process/KRR formulation, molecular covariance is decomposed into atomic covariances,

$$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$

with a Gaussian kernel and atom-type matching. This yields a transferable mapping from local atomic environments to property contributions rather than a monolithic molecular map (Huang et al., 2017).

The same paper defines amons as atoms-in-molecules-based fragments selected on-the-fly from a query system. The fragment-selection rule is graph-theoretic and chemically constrained: connected subgraphs are enumerated, the original hybridization and valence state must be preserved, and exact subgraph isomorphism to the query graph is required. For noncovalent systems the procedure is adapted with vdW bonds and special selection rules, while for very large biomolecules static amons are obtained as exact cutouts around a target atom within a cutoff radius. This suggests that one early route to universality was not a single end-to-end universal potential, but an atomwise reconstruction strategy based on a finite dictionary of recurring local environments (Huang et al., 2017).

A complementary precursor is the Gaussian multipole (GMP) featurization, which was introduced to avoid element-specific feature growth. GMP encodes chemical identity through an approximate electron density built from atom-centered Gaussians and probes that density with radial Gaussians and Maxwell–Cartesian spherical harmonics. The basic feature is

$$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$

and rotational invariance is imposed through grouped norms over permutation-related harmonics. The resulting feature vector has fixed dimension regardless of the number of elements present, enabling a single neural network rather than separate subnetworks per element type. The authors explicitly position this as a route toward a more universal atomistic model whose input representation can be reused across many element types, system classes, and datasets (Lei et al., 2021).

2. UMA architecture, model family, and scaling laws

In its current form, UMA is built on eSEN, an equivariant graph neural network, with the principal architectural novelty given by the Mixture of Linear Experts (MoLE) mechanism. The standard expert combination is

$y=\sum_k \alpha_k (W_k x),$

which UMA rewrites as

$y=W^*x,\qquad W^*=\sum_k \alpha_k W_k.$

Because the expert mixture is computed once from global, time-invariant information and merged before the forward pass, the model gains large total parameter count without a proportional inference penalty. The inputs therefore include not only atom positions and atomic numbers, but also global system descriptors such as charge, spin multiplicity, and the DFT task label (Wood et al., 30 Jun 2025).

The published UMA family contains three principal variants. UMA-S has 150M total parameters, 6M active parameters, about 16 inferences/sec for 1k atoms, and fits 100k+ atoms on an 80GB GPU. UMA-M has 1.4B total parameters, 50M active parameters, about 3 inferences/sec for 1k atoms, and fits 10k+ atoms on an 80GB GPU. UMA-L has 700M total parameters, 700M active parameters, about 1.6 inferences/sec for 1k atoms, and fits 1k+ atoms on an 80GB GPU. On a typical H100 benchmark with approximately 50 neighbors per atom, UMA-S can simulate 1000 atoms at 16 steps/sec, corresponding to about 1.4 ns/day (Wood et al., 30 Jun 2025).

A central technical contribution is the empirical study of scaling laws at atomistic-model scale. Training compute is modeled as

$C(N,D)\approx \kappa N D,$

with $\kappa\approx 270$ FLOPs/parameter/atom for the UMA-M-like setting. Compute-optimal size laws are then fit as

$$\log(N^*(C))=\alpha\log(C)+A,\qquad \log(D^*(C))=\beta\log(C)+B,$$

and the usual two-variable loss ansatz is written as

$$\tilde{L}(N,D)=\hat{E}+\frac{\hat{A}}{N^{\hat{\alpha}}}+\frac{\hat{B}}{D^{\hat{\beta}}}.$$

The reported fit coefficients are Dense: $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$0 and $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$1; MoLE: $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$2 and $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$3. The authors conclude that model capacity must keep growing with dataset size, and that MoLE can reduce the number of active parameters needed for the same loss by roughly a factor of $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$4 for UMA-M (Wood et al., 30 Jun 2025).

Training is organized as a two-stage schedule: a direct-force pretraining stage followed by a conservative finetuning stage in which the force head is removed and forces and stresses are computed by autograd to enforce energy conservation. The implementation uses BF16 pretraining, FP32 finetuning, max-atom batching, reduced neighbor counts during pretraining, graph parallelism, fully-sharded data parallelism for large MoLE layers, activation checkpointing, and training runs up to about 10B total parameters. BF16 alone is reported to degrade accuracy by 20–50% depending on task (Wood et al., 30 Jun 2025).

3. Training data, data harmonization, and universal dataset design

The combined dataset used for UMA contains 459,156,663 training examples and over 30 billion atoms. The five constituent datasets are OMat24 with 100,824,585 materials structures, OMol25 with 75,889,983 molecular structures, OC20++ with 229,054,043 catalysis structures, OMC25 with 24,870,226 molecular crystals, and ODAC25 with 28,517,826 MOF / DAC structures. Average structure size ranges from 19 atoms in OMat24 to 178 atoms in ODAC25, with a maximum of 350 atoms, and the combined pairwise element interactions cover nearly the entire periodic table except radioactive elements (Wood et al., 30 Jun 2025).

A major obstacle for universal MLIPs is that available datasets are generated under different exchange-correlation functionals, basis sets, core-electron treatments, boundary conditions, and software packages. Total Energy Alignment (TEA) was introduced to address this directly. TEA has two stages: Inner Core Energy Alignment (ICEA) and Atomization Energy Correction (AEC). If two methods have the same atomization energy for a given geometry, total energies can be shifted from one energy scale to another using isolated atom energies; residual fidelity differences are then modeled with a linear scaling

$$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$5

which induces the same scaling on forces,

$$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$6

Using TEA, the authors trained MACE-Osaka24, described as the first open-source neural network potential model based on a unified dataset covering both molecular and crystalline systems (Shiota et al., 2024).

A different but closely related data philosophy appears in the MAD dataset, which was designed specifically to train universal models that can provide reasonable predictions for arbitrary structures. MAD contains 95,595 structures and 85 elements with atomic numbers ranging from 1 to 86, excluding astatine, and deliberately prioritizes massive atomic diversity over physical plausibility. Starting from stable seed structures, the dataset is expanded through aggressive perturbations such as MC3D-rattled, MC3D-random, MC3D-surface, and MC3D-cluster constructions. All labels are computed with a single consistent DFT protocol based on Quantum ESPRESSO v7.2, AiiDA, PBEsol, no spin polarization, SSSP v1.2 efficiency pseudopotentials, a 110 Ry plane-wave cutoff, a 1320 Ry charge-density cutoff, Marzari-Vanderbilt-De Vita-Payne cold smearing with spread 0.01 Ry, and a Γ-centered k-point spacing of 0.125 Å⁻¹ in periodic directions. This compact dataset is reported to enable training universal interatomic potentials competitive with models trained on traditional datasets with two to three orders of magnitude more structures (Mazitov et al., 24 Jun 2025).

These two adjacent developments clarify that universality in atomistic ML is not only an architectural question. One route standardizes scale by massive cross-domain training; another aligns heterogeneous corpora almost without redundant calculations; a third deliberately broadens the support of configuration space through aggressive distortion. A plausible implication is that UMA-scale models depend simultaneously on model capacity, dataset breadth, and coherent reference labeling (Wood et al., 30 Jun 2025).

4. Cross-domain empirical performance

UMA is evaluated as a zero-shot or near-zero-shot universal model across materials, catalysis, molecules, molecular crystals, and MOFs. On Matbench Discovery, UMA-M achieves F1 = 0.930, reported as the best to date. On OC20 OOD-both, the adsorption-energy MAE is 70.2 meV for UMA-S, 46.5 meV for UMA-M, and 43.5 meV for UMA-L, compared with 306.5 meV for eqV2-OC20 and 343.3 meV for GemNet-OC20. On AdsorbML, success rate rises to 71.12% for UMA-M and 74.41% for UMA-L, versus 60.80% for eqV2-OC20 and 54.88% for GemNet-OC20. The paper highlights about a 25% improvement in success rate for UMA-L (Wood et al., 30 Jun 2025).

On the molecular domain, the challenging OMol25 test splits show that UMA improves over a domain-specific eSEN baseline. On the test OOD-Comp split, UMA-M reaches 0.42 meV/atom energy MAE and 3.89 meV/Å force MAE; UMA-L improves further to 0.34 meV/atom and 2.92 meV/Å. In ligand-strain prediction, UMA-M reaches 2.45 meV MAE, which the authors note is at or below chemical accuracy. The same evaluation suite includes protein-ligand interaction, ionization/affinity, spin-gap, and distance-scaling benchmarks (Wood et al., 30 Jun 2025).

On OMC25 molecular-crystal data, UMA-M obtains 0.8 meV/atom energy MAE and 3.0 meV/Å force MAE, while UMA-L reaches 0.6 meV/atom and 2.3 meV/Å. On the 7th CCDC CSP blind test polymorph subset, UMA-L achieves 2.49 kJ/mol lattice-energy MAE and 1.00 match rate. On the ODAC adsorption benchmark for MOFs, UMA-L reaches 291.1 meV adsorption-energy MAE and 6.5 meV/Å force MAE, compared with 316.0 meV and 7.2 meV/Å for eqV2-ODAC. These results are repeatedly framed as evidence that one model family can now match or exceed many specialized baselines across chemistry and materials tasks (Wood et al., 30 Jun 2025).

Related universal models derived from different training strategies show comparable broad-domain behavior. MACE-Osaka24-large, trained on TEA-aligned molecular and crystalline data, attains 0.457 kcal/mol MAE on 78 drug-like biaryl torsions, 0.404 eV MAE on Transition1x reaction energy, 0.265 eV MAE on Transition1x energy barrier, and 0.018 Å overall MAE on bulk-crystal lattice constants. The same study argues that integrating molecular data does not significantly degrade inorganic performance (Shiota et al., 2024). This suggests that cross-domain transfer is not exclusive to one architecture, but UMA distinguishes itself by combining that transfer with very large-scale pretraining and a speed-preserving MoLE design.

5. Fine-tuning and application workflows

Although UMA is presented primarily as a zero-shot universal model, the literature also treats it as a foundation model that can be adapted to narrower domains. In oxygen-plasma interactions with multilayer WS$$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$7, the specific checkpoint used is uma-s-1p1, the UMA-S variant under the Open Catalyst 2020 (OC20) task. The adaptation procedure is an iterative loop: MD exploration with the current model, configuration sampling, DFT labeling, fine-tuning, and evaluation on newly generated configurations. Diversity selection uses SOAP, PCA, and FPS, with SOAP features reduced to 50 principal components capturing about 99% of the variance, followed by Farthest Point Sampling. Labels are generated by periodic DFT in Quantum ESPRESSO at PBE + D3 + $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$8 + spin, with Hubbard $$E^I-\varepsilon_0^I=\phi(\mathbf{M}^I)^\top \mathbf{w}+\epsilon.$$9 correction of 2.87 eV on W 5d orbitals (Kwon et al., 19 Jun 2026).

The training targets in that study are energy, force, and stress, with loss

$$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$0

The force channel uses Huber loss rather than MSE, while energy and stress use MSE. The production results use layer freezing, with 90/10 train/validation split, 55 epochs, cosine learning-rate schedule, 5-epoch linear warmup, and EMA weights. Even without fine-tuning, the pretrained model reproduces chemisorbed S and O coverage under 15 eV O$$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$1 and O$$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$2 bombardment. After iterative fine-tuning, the best reported model, R3, reaches Energy MAE = $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$3 eV/atom, Force MAE = 0.076 eV/Å, Stress MAE = 0.034 GPa, with $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$4 for energy, $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$5 for forces, and $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$6 for stress (Kwon et al., 19 Jun 2026).

A second application is molecular crystal structure prediction (CSP). In FastCSP, UMA is the MLIP core of a fully open, high-throughput workflow that combines random structure generation using Genarris 3.0 with relaxation and free-energy calculations powered entirely by UMA. The workflow generates random crystal packings across compatible space groups and multiple $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$7 values, compresses them with Rigid Press, deduplicates them using Pymatgen StructureMatcher, relaxes them with UMA-Small v1.1, deduplicates again, and ranks by 0 K lattice energy or, optionally, by Helmholtz free energy or Gibbs free energy. The thermodynamic ranking uses

$$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$8

with the harmonic approximation for $$K_{ij}=\sum_I\sum_J K_{ij}^{IJ}, \qquad K_{ij}^{IJ}=k(\mathbf{M}_i^I,\mathbf{M}_j^J),$$9 and a Vinet equation of state in the quasiharmonic approximation (Gharakhanyan et al., 4 Aug 2025).

The benchmark set in FastCSP contains 28 mostly rigid molecules and 36 crystal structures. The workflow is reported to have generated the experimental structure for all 28 molecules, with one caveat: CILJIQ Form II required increasing generation from 500 to 1000 structures per space group. Across the benchmark, 17 of 28 target molecules had an experimental structure or known polymorph ranked as the lattice-energy global minimum, 8 more were ranked within the top 4, all experimental structures were within 5 kJ/mol per molecule of the global minimum, and the workflow achieved over 94% recall within the top 10 lowest-energy predictions and within 3 kJ/mol of the global minimum. Comparing UMA-relaxed structures to PBE-D3, the reported metrics are 1.16 kJ/mol MAE for relaxation energies, 0.94 Spearman rank correlation, 0.22 Å average RMSD$$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$0, and 90% average match rate. Runtime is approximately 15 seconds per relaxation on an NVIDIA H100 80 GB GPU, and a single CSP run can be completed within hours on tens of modern GPUs (Gharakhanyan et al., 4 Aug 2025).

6. Uncertainty, reliability, and limitations

A central problem for universal atomistic foundation models is how to determine when predictions are trustworthy without running new DFT calculations. A recent answer is the uncertainty metric $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$1 derived from a heterogeneous ensemble of pretrained uMLIPs. The recommended metric is the inverse-RMSE-weighted formulation

$$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$2

with model weights

$$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$3

The optimal ensemble size is 11 models. On OMat24, the reported correlations are $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$4 for equal-weight $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$5, $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$6 for weighted $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$7, and $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$8 for weighted-average $$\mu_{i,abc}=\langle S_{abc}\times G_i,\hat{\rho}\rangle,$$9; the final recommended metric is $y=\sum_k \alpha_k (W_k x),$0 (Liu et al., 28 Jul 2025).

The same study shows that uncertainty can drive filtering and data selection. On OMat24, using $y=\sum_k \alpha_k (W_k x),$1, energy RMSE stays below $y=\sum_k \alpha_k (W_k x),$2 and force RMSE below $y=\sum_k \alpha_k (W_k x),$3 up to about 80% coverage, whereas Orb-confidence reaches comparable accuracy only below roughly 25% coverage for energy and 40% for force. A practical threshold

$y=\sum_k \alpha_k (W_k x),$4

is suggested as a rule of thumb for configurations that can be predicted with RMSE $y=\sum_k \alpha_k (W_k x),$5, while $y=\sum_k \alpha_k (W_k x),$6 yields about $y=\sum_k \alpha_k (W_k x),$7 RMSE. The uncertainty-aware model distillation framework then uses teacher predictions when $y=\sum_k \alpha_k (W_k x),$8 and DFT labels when $y=\sum_k \alpha_k (W_k x),$9; for tungsten, accuracy comparable to full-DFT training is achieved using only about 4% DFT-labeled configurations, and for MoNbTaW no additional DFT calculations are required (Liu et al., 28 Jul 2025).

Universality, however, is explicitly bounded. In the AML/amons framework, the locality assumption is not exact for strong nonlocal physics such as highly correlated phenomena, Peierls or Jahn–Teller distortions, and metal–insulator transitions; conjugated polymers require larger amons as the query length grows, and proteins or long-range interactions require static or cutoff-based fragments rather than fully relaxed ones (Huang et al., 2017). In TEA-based universal models, the method depends on suitable isolated-atom reference energies, stronger correlation or relativistic systems may be harder to align reliably, a single global scaling factor may be too simple for some cases, and a 4.5 Å cutoff limits some BCC lattice predictions (Shiota et al., 2024). In MAD, consistency is achieved partly by simplifying or neglecting magnetism, electron correlation, and dispersion, and convergence falls to about 55% for the deliberately random MC3D-random subset (Mazitov et al., 24 Jun 2025).

The modern UMA family is also constrained by the reference tasks on which it is trained. Because the same structure can have different labels under different DFT settings, UMA includes the DFT task label as an input rather than presuming a single universal energy scale (Wood et al., 30 Jun 2025). Application studies reinforce the same point. In plasma–WS$y=W^*x,\qquad W^*=\sum_k \alpha_k W_k.$0 chemistry, the pretrained OC20 checkpoint already captures the main production observables, but fine-tuning is motivated precisely because spin polarization and Hubbard $y=W^*x,\qquad W^*=\sum_k \alpha_k W_k.$1 matter for W–S–O chemistry (Kwon et al., 19 Jun 2026). In FastCSP, the benchmark is restricted to mostly rigid molecules, excludes $y=W^*x,\qquad W^*=\sum_k \alpha_k W_k.$2, cocrystals, salts, hydrates, and solvates, and identifies Target XVI as a coverage-dependent failure due to unusual diazide-carbonyl interactions underrepresented in training data (Gharakhanyan et al., 4 Aug 2025). A common misconception is therefore that “universal” means exact or reference-independent across all chemistry. The literature instead supports a narrower interpretation: universality is practical, data-driven, and strongly conditioned on coverage, labeling conventions, and the target observable.