Universal Models for Atoms
“When you keep improving technology, eventually you reach a point where you have to improve the materials.”
That was what Ekin Doğuş Çubuk, who would later co-found Periodic Labs, said when DeepMind unveiled GNoME in 2023. The more you think about it, the more it rings true. Longer-lasting batteries, faster semiconductors, superconductors that work closer to room temperature—much of what engineering aspires to today runs up against the same question: “Does a material with those properties even exist?”
The problem is that finding materials remains painfully slow. Once you multiply possible combinations of elements by their ratios and crystal structures, the number of candidates is effectively infinite. Yet researchers must find what they need through a handful of experiments within the narrow range that experience points them toward. Density functional theory (DFT), a standard tool in computational chemistry, helps to some extent. But because its computational cost grows roughly with the cube of the number of atoms, it can handle only a few hundred atoms over fleeting timescales.
UMA—Universal Models for Atoms—released by Meta FAIR in 2025, takes aim at precisely this bottleneck. In one sentence: it brings the foundation-model approach that worked for LLMs into the atomic world—and shows that it actually works there, too.
500 Million Atomic Structures in a Single Model
UMA is a machine-learning interatomic potential (MLIP) that predicts the energies and forces of atomic structures in place of DFT. MLIPs themselves are not new. What sets UMA apart is its scale and versatility. Meta combined datasets it had built over the previous five years, covering molecules (OMol25), materials (OMat24), catalysts (OC20), molecular crystals (OMC25), and metal-organic frameworks, or MOFs (ODAC25). The resulting training data encompassed 500 million atomic systems and more than 30 billion atoms—the largest training effort in the field to date. The single model trained on this data covers nearly the entire chemical space, excluding radioactive elements.
Until now, MLIPs have largely operated in silos. A model trained on materials cannot handle molecules; a molecular model falls apart when confronted with a surface. Yet the problems we most want to solve often lie precisely at those boundaries. Batteries involve interfaces between electrodes—inorganic materials—and electrolytes—organic molecules. Catalysis involves interactions between molecules and surfaces. Carbon capture involves gas molecules binding to MOF frameworks. With models divided by domain, even posing these questions is difficult.
Scaling Up Without Slowing Down: MoLE
Of course, larger datasets call for larger models, and larger models tend to run more slowly. The result can be a model that is accurate but impractical. UMA’s answer is MoLE: Mixture of Linear Experts. It borrows the idea of mixture-of-experts architectures from LLMs, but adapts it to atomic simulations. Because an energy surface must be smooth, the experts are linear, avoiding discontinuities while preserving rotational equivariance.
The crucial detail is that routing—the decision about how much to use each expert—depends only on information that remains constant throughout a simulation, such as elemental composition, charge, and spin. This means the expert weights can be combined into a single set before the simulation begins. As a result, UMA-M has 1.4 billion parameters in total, but uses only about 50 million effective parameters for each structure, allowing it to run molecular dynamics (MD) simulations efficiently over millions of steps.
What the Numbers Tell Us
The results are impressive. Without any fine-tuning, a single model was evaluated on benchmarks spanning materials, catalysts, molecules, molecular crystals, and MOFs—and matched or outperformed models specialized for each field. It achieved the highest F1 score on Matbench Discovery, a leading benchmark for predicting material stability; improved the success rate of adsorption-energy searches in AdsorbML by 25%; and reached chemical accuracy for ligand strain energies, a quantity used in drug design.
What about speed? On a single GPU, UMA-S simulates a 1,000-atom system at 16 steps per second, equivalent to 1.4 nanoseconds per day, and can fit systems of more than 100,000 atoms into memory. These are system sizes and simulation timescales that would be out of reach for DFT.
Another finding deserves attention. Models trained on a single domain suffered from overfitting, whereas models trained across multiple domains generally converged to lower errors. Molecular data also helped the model learn about molecular crystals. This suggests that different areas of chemistry can share a common representation space—the same pattern we saw when LLMs were trained on mixtures of languages and tasks.
So What Changes?
A speedup of several thousand times does more than make things convenient. It changes the kinds of science we can do. A search that once relied on intuition to narrow the field to a few hundred candidates can become a computational screening process covering millions. Unconventional candidates beyond the reach of human intuition get a chance to survive. Instead of selecting materials solely from static structures, researchers can also examine dynamic mechanisms such as interfacial reactions, ion transport, and defect motion.
There is a natural connection here to autonomous laboratories such as Periodic Labs. No matter how thoroughly robots automate synthesis and measurement, the whole loop remains slow if the “brain” deciding what to try next is slow. A general-purpose potential capable of running inference dozens of times per second could serve as the computational brain that generates and filters hypotheses in real time. One of the core components of a self-driving lab—where prediction, synthesis, measurement, and learning proceed without human intervention—has now arrived, with its model weights publicly available.
What Still Needs Work
None of this means UMA has solved automated materials discovery. Its 6 Å cutoff prevents it from accounting for interactions between distant atoms, while charge and spin are handled through discrete embeddings that do not generalize to values absent from the training data. Its accuracy is still bounded by DFT, its teacher: where DFT gets things wrong, UMA inherits those errors. Follow-up evaluations have also reported uneven performance in less well-covered regions of chemical space, such as transition-metal complexes.
Above all, the gap between “a material that looks promising in a calculation” and “a material we can actually make in the lab”—the problem of synthesizability—remains open.
