Universal Model for Atoms Overview

• Universal Model for Atoms is a unified framework that employs equivariant graph neural networks to predict interatomic energies, forces, and stresses across chemically diverse systems.
• It integrates a Mixture-of-Linear-Experts mechanism with global task embeddings to enable zero-shot transfer across domains without domain-specific fine-tuning.
• Empirical scaling laws and benchmark evaluations demonstrate UMA's efficiency and scalability in applications like crystal structure prediction and interatomic potential modeling.

Universal Model for Atoms, usually abbreviated UMA, denotes a family of equivariant graph neural network interatomic potentials trained on a single combined corpus of $\sim 459$ million DFT calculations, or $\sim 30$ billion atoms total, spanning materials, molecules, catalysts, molecular crystals, and metal-organic frameworks (Wood et al., 30 Jun 2025). Its central claim is that a single multi-task model, conditioned on atomic structure and global descriptors such as total charge, spin multiplicity, and the DFT task to emulate, can compute atomistic energies, forces, and stresses across chemically distinct domains without domain-specific fine-tuning (Wood et al., 30 Jun 2025). The term “universal” also appears in a different sense in few-body atomic physics, where it refers to van der Waals or Efimov parameter reduction rather than cross-domain machine-learning transfer (Wang et al., 2014, Kievsky et al., 2017).

1. Architectural formulation

UMA builds on an equivariant graph neural network in which node embeddings are constructed from per-atom spherical harmonics up to degree $L_{\max}$, initialized by atomic number and augmented by global embeddings for total charge, spin multiplicity, and a one-hot “task” vector identifying which DFT setting to emulate. Message passing is implemented through $T$ layers of SO(2) convolutions, each followed by node-wise feed-forward blocks and layer normalization. A single node-wise block predicts total energy, while forces and, optionally, stress are obtained either by direct heads during pre-training or by auto-gradients during fine-tuning (Wood et al., 30 Jun 2025).

The distinctive UMA mechanism is the Mixture-of-Linear-Experts (MoLE). Instead of a single linear map $W$, each linear block uses $K$ experts $\{W_k\}_{k=1}^K$ combined through dense gating coefficients $\{\alpha_k\}$: $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$ Because the gating coefficients depend only on global features such as element composition, charge, spin, and DFT task, UMA can precompute

$W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$

thereby preserving rotational equivariance and adding no extra cost at inference. The gating network is a 3-layer MLP plus softmax acting on the concatenated global embeddings (Wood et al., 30 Jun 2025).

In downstream documentation, the FastCSP deployment describes UMA as a message-passing, equivariant graph neural network that predicts total potential energy, atomic forces, and stress tensor from an atomic configuration $\sim 30$0, with

$\sim 30$1

That account states that the practical implementation uses the eSEN architecture, where the neighbor interaction $\sim 30$2 is realized by a self-attention module within a cutoff $\sim 30$3, node updates and readouts are built from gated equivariant layers, and a MoLE head is used in the final readout to scale model capacity while preserving fast inference (Gharakhanyan et al., 4 Aug 2025).

2. Training corpus, preprocessing, and optimization

UMA is trained on a single combined dataset of approximately half a billion DFT calculations assembled from five public benchmarks. The constituent datasets are OMat24 with 100 million samples over 89 elements using PBE/VASP and sampling weight $\sim 30$4; OMol25 with 76 million samples over 83 elements using $\sim 30$5B97M-V/ORCA and weight $\sim 30$6; OC20++ with 229 million samples over 56 elements using RPBE/VASP and weight $\sim 30$7; OMC25 with 25 million samples over 12 elements using PBE-D3/VASP and weight $\sim 30$8; and ODAC25 with 29 million samples over 70 elements using PBE-D3/VASP and weight $\sim 30$9 (Wood et al., 30 Jun 2025).

To merge these chemically and methodologically distinct sources into a single multi-task model, UMA introduces a global task embedding, applies a heat-of-formation referencing of the total energy,

$L_{\max}$0

normalizes all labels by their force-RMS, and samples from each dataset in proportion to its assigned weight (Wood et al., 30 Jun 2025). The need for an explicit task embedding is a reminder that UMA is universal across several electronic-structure domains, but not agnostic to the reference method.

Training follows a two-stage schedule. In the pre-training stage, UMA uses BF16 precision, a direct force head without autograd, radius cutoff $L_{\max}$1 Å, max-neighbors $L_{\max}$2, AdamW optimization, cosine learning-rate schedule with peak $L_{\max}$3, warm-up $L_{\max}$4 of steps, and weight decay $L_{\max}$5. The number of pre-train steps is $L_{\max}$6M for UMA-S, $L_{\max}$7M for UMA-M, and $L_{\max}$8M for UMA-L. The loss is

$L_{\max}$9

with $T$0 and $T$1, except $T$2 for OMol25 energy (Wood et al., 30 Jun 2025).

In the fine-tuning stage, UMA switches to FP32 precision, uses autograd for conservative forces and stress, increases max-neighbors to $T$3 while keeping the same radius, applies a cosine schedule with peak learning rate $T$4, warm-up $T$5, and weight decay $T$6, and runs $T$7M steps for UMA-S, $T$8M for UMA-M, and $T$9M for UMA-L. The fine-tuning objective is

$W$0

with $W$1, $W$2, and $W$3 (Wood et al., 30 Jun 2025).

3. Scaling laws and computational profile

A notable feature of UMA is that it is accompanied by explicit empirical scaling laws. Training compute is approximated as

$W$4

with $W$5 FLOPs per parameter per atom for UMA-M. On the compute-optimal frontier, the optimal model size and dataset size satisfy

$W$6

The fitted exponents are $W$7 for dense models and $W$8 for MoLE models (Wood et al., 30 Jun 2025).

Restricting to that frontier, the validation loss follows

$W$9

with $K$0 for dense models and $K$1 for MoLE models. The corresponding interpretation given in the paper is that an MoLE model can match the loss of a dense model with approximately $K$2 fewer active parameters (Wood et al., 30 Jun 2025).

The three released sizes illustrate how this trade-off is operationalized. UMA-S has $K$3M total parameters with $K$4M active parameters and $K$5 experts; UMA-M has $K$6B total parameters with $K$7M active parameters and $K$8 experts; UMA-L has $K$9M parameters and is dense. Inference measurements, excluding graph-construction overhead, are reported on a single H100 GPU for a periodic system with $\{W_k\}_{k=1}^K$0 neighbors per atom in TF32. At $\{W_k\}_{k=1}^K$1 atoms the speeds are $\{W_k\}_{k=1}^K$2, $\{W_k\}_{k=1}^K$3, and $\{W_k\}_{k=1}^K$4 steps per second for UMA-S, UMA-M, and UMA-L, respectively; at $\{W_k\}_{k=1}^K$5 atoms they are $\{W_k\}_{k=1}^K$6, $\{W_k\}_{k=1}^K$7, and OOM; and at $\{W_k\}_{k=1}^K$8 atoms UMA-S reaches $\{W_k\}_{k=1}^K$9 steps per second while the larger models are OOM (Wood et al., 30 Jun 2025). This profile explains why the “universal” claim is tied not only to accuracy but also to the feasibility of running one model across many atomistic workloads.

4. Zero-shot evaluation across domains

UMA is evaluated in a zero-shot setting, meaning without domain-specific fine-tuning, across benchmarks in materials, catalysis, molecules, molecular crystals, and metal-organic frameworks. On Matbench Discovery, higher F1 is better, and the reported values are $\{\alpha_k\}$0 for UMA-S, $\{\alpha_k\}$1 for UMA-M, and $\{\alpha_k\}$2 for UMA-L, compared with $\{\alpha_k\}$3 for eSEN-30M. On OC20 S2EF adsorption, UMA-M reports adsorption-energy errors of $\{\alpha_k\}$4 meV ID and $\{\alpha_k\}$5 meV OOD, with force MAE $\{\alpha_k\}$6 meV ID and $\{\alpha_k\}$7 meV OOD; the comparison entry eV$\{\alpha_k\}$8-OC20 reports $\{\alpha_k\}$9, $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$0, $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$1, and $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$2 meV, respectively (Wood et al., 30 Jun 2025).

On AdsorbML, the global-minima success rates are $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$3, $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$4, and $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$5 for UMA-S, UMA-M, and UMA-L, compared with $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$6 for the eV$$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$7 baseline. On OMol25 OOD-Comp, the energy error $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$8 is $$y=\sum_{k=1}^K \alpha_k\,(W_k x), \qquad \sum_k \alpha_k=1.$$9, $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$0, and $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$1 meV/atom for UMA-S, UMA-M, and UMA-L, compared with $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$2 meV/atom for eSEN-OMol. On OMC25 crystal lattices, $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$3 is $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$4, $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$5, and $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$6 meV/atom for UMA-S, UMA-M, and UMA-L, compared with $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$7 meV/atom for eSEN-OMC. On ODAC25 MOF adsorption, the OOD adsorption-energy errors are $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$8, $W^*=\sum_{k=1}^K \alpha_k W_k, \qquad y=W^*x,$9, and $\sim 30$00 meV for UMA-S, UMA-M, and UMA-L, compared with $\sim 30$01 meV for eqV2-ODAC (Wood et al., 30 Jun 2025).

These results underwrite the paper’s central contention that a single, unfine-tuned model can perform similarly or better than specialist models across heterogeneous chemical domains. They also clarify that “universality” in UMA is empirical rather than axiomatic: it is established by cross-benchmark transfer under shared training and inference rules, not by a proof of method-independence.

5. Molecular-crystal structure prediction with FastCSP

A concrete downstream realization of UMA is FastCSP, an open-source, high-throughput crystal-structure-prediction workflow for molecular crystals. Its workflow proceeds through six stages: random-structure generation via Genarris 3.0 using $\sim 30$02 structures per space group and up to $\sim 30$03; pre-filtering by rigid-press compression and pymatgen deduplication; full geometry and lattice relaxation with UMA, discarding failures or connectivity changes; second deduplication with retention of structures within $\sim 30$04 kJ/mol of the global minimum; optional finite-temperature free-energy evaluation with UMA through harmonic approximation phonons in Phonopy and quasi-harmonic approximation using a Vinet EOS and Legendre transform to obtain $\sim 30$05; and final ranking by UMA lattice energy or free energy at the desired $\sim 30$06 (Gharakhanyan et al., 4 Aug 2025).

In that workflow, the UMA-S-1.1 “Small” model is benchmarked against PBE-D3 on $\sim 30$07 mostly rigid molecules covering $\sim 30$08 experimental polymorphs. The reported potential-energy-surface fidelity is a relative lattice-energy MAE of $\sim 30$09 kJ/mol per molecule with Spearman rank $\sim 30$10. The reported geometry-relaxation agreement is average $\sim 30$11 Å, with $\sim 30$12 of UMA-relaxed structures matching PBE-D3 topology under $\sim 30$13 Å. In terms of CSP recall, $\sim 30$14 of known polymorphs lie within the top-10 UMA-ranked minima at $\sim 30$15 K lattice energy, and all known forms are within $\sim 30$16 kJ/mol of the global minimum according to UMA. On an NVIDIA H100 with $\sim 30$17 GB memory, a full relaxation of one crystal takes approximately $\sim 30$18 s (Gharakhanyan et al., 4 Aug 2025).

FastCSP is also where several practical limits of the UMA paradigm are made explicit. The manuscript notes larger errors for chemistry not well represented in OMC25, exemplified by diazide–carbonyl interactions in Target XVI; it states that the workflow has not yet been tested on molecules with extensive conformational flexibility $\sim 30$19, co-crystals, salts, hydrates, or solvates; and it remarks that final ranking of $\sim 30$20 differences below $\sim 30$21 kJ/mol may still require higher-level DFT such as PBE0+MBD for some systems, including glycine polymorphs (Gharakhanyan et al., 4 Aug 2025). Thus, the crystal-structure-prediction application presents UMA as a high-accuracy production model, but not as a complete replacement for all higher-level electronic-structure verification.

6. Broader scope and distinct meanings of “universality”

Later work in biomolecular simulation places UMA within a broader landscape of atomic foundation models. UBio-MolFM introduces UBio-Mol26, a $\sim 30$22-million-frame training set, E2Former-V2 with Equivariant Axis-Aligned Sparsification and Long-Short Range modeling, and a Three-Stage Curriculum Learning protocol for large bio-systems. In its large-system benchmarks, E2Former-V2 reports approximately $\sim 30$23 higher inference throughput than state-of-the-art equivariant models, specifically UMA-S and MACE, on $\sim 30$24–$\sim 30$25K atom systems, while targeting ab initio-level fidelity for systems up to approximately $\sim 30$26 atoms (Huang et al., 13 Feb 2026). A plausible implication is that UMA has become a reference point against which later domain-specialized universal atomic models are measured.

A different research tradition uses “universal” in the sense of low-energy few-body physics. In the universal van der Waals model for three ultracold atoms, the entire two-spin, two-channel Feshbach resonance is characterized by the two dimensionless two-body parameters $\sim 30$27 and $\sim 30$28, together with the van der Waals length $\sim 30$29. The three-body Schrödinger equation is solved in hyperspherical coordinates with pairwise Lennard-Jones $\sim 30$30–$\sim 30$31 potentials, and the resulting theory predicts three-body recombination and atom-dimer collision rates without adjustable fitting parameters. For ultracold Cs near $\sim 30$32 G, using $\sim 30$33 and $\sim 30$34, the model reproduces the first Efimov recombination resonance at $\sim 30$35, the double-minimum structure in recombination at small $\sim 30$36, and the atom-dimer loss resonance at $\sim 30$37 (Wang et al., 2014).

Near the unitary limit in three-boson systems, universality is expressed through the angle $\sim 30$38 defined by $\sim 30$39, the three-body parameter $\sim 30$40, and Efimov relations such as

$\sim 30$41

for the ground state. In that framework, the atom-dimer scattering length obeys

$\sim 30$42

with $\sim 30$43, $\sim 30$44, and $\sim 30$45, and finite-range corrections are studied with local and non-local Gaussian potentials fixed to reproduce $\sim 30$46. For three $\sim 30$47He atoms, mapping a Gaussian model to the LM2M2 potential gives $\sim 30$48, compared with direct LM2M2 calculations of approximately $\sim 30$49 (Kievsky et al., 2017).

This distinction suggests that the same adjective labels two different scientific programs. In current computational chemistry and materials science, “Universal Model for Atoms” refers to a trained cross-domain interatomic-potential family such as UMA (Wood et al., 30 Jun 2025). In AMO few-body physics, “universal” refers instead to the compression of observables into a small set of low-energy scales and dimensionless parameters (Wang et al., 2014, Kievsky et al., 2017).