Universal Interatomic Potentials

• Universal Interatomic Potentials are comprehensive models that predict energies, forces, and stresses using machine-learning and analytic approaches without material-specific retraining.
• They leverage advanced graph neural network architectures and physics-aware formalisms to efficiently model equilibrium properties, though challenges persist in high-temperature and reactive regimes.
• UIP applications span phonons, elastic properties, surfaces, defects, and framework materials, enabling high-throughput screening and informed fine-tuning for improved accuracy.

Universal Interatomic Potentials (UIPs) are interatomic models intended to operate across a wide chemical space without retraining or material-specific parameterization. In recent benchmark literature, the term is used primarily for universal machine-learning interatomic potentials (uMLIPs): pre-trained models that predict energies, forces, and stresses for diverse structures at greatly reduced computational cost relative to density functional theory (DFT), while a broader usage also includes universal analytic interatomic potentials such as UFF, Dreiding, and GFN-FF. Across recent studies, UIPs have been assessed on phonons, thermodynamics, elastic properties, defects, surfaces, zeolites, metal-organic frameworks (MOFs), 2D materials, and perovskite oxides, with strong performance in many equilibrium and near-equilibrium settings but persistent limitations in out-of-domain, reactive, high-temperature, and far-from-equilibrium regimes (Yu et al., 2024, Ito et al., 9 Sep 2025, Edwards et al., 28 Apr 2026).

1. Definition, scope, and major model classes

In systematic assessments, “universal” has been defined as applicability across a wide chemical space—potentially the entire periodic table—without retraining or fine-tuning for each specific material or composition (Yu et al., 2024). Within this definition, recent benchmark papers have treated M3GNet, CHGNet, MACE, ALIGNN, SevenNet, MatterSim, ORB, OMat24, and related models as representative uMLIPs, with most architectures based on graph neural networks, equivariant message passing, or closely related graph-transformer designs (Loew et al., 2024).

The same literature also shows that UIP is not restricted to machine learning. In zeolite benchmarking, universal analytic interatomic potentials—UFF, Dreiding, and GFN-FF—were evaluated alongside pretrained universal MLIPs such as CHGNet, ORB-v3, MatterSim, eSEN-30M-OAM, PFP-v7, and EquiformerV2-lE4-lF100-S2EFS-OC22 (Ito et al., 9 Sep 2025). This distinction is substantive rather than terminological: analytic UIPs rely on fixed functional forms and broad parameter sets, whereas uMLIPs learn the potential-energy surface from large DFT-derived datasets.

A second defining feature of current uMLIPs is scale. In one phonon benchmark, the listed training data sizes ranged from 188K structures for M3GNet to 110M for OMat24, with intermediate regimes such as 1.6M for CHGNet, MACE, SevenNet, and ORB, and 17M for MatterSim (Loew et al., 2024). This broadening of training coverage underlies the claim of universality, but it does not eliminate the difference between interpolation across diverse known environments and extrapolation to chemically or structurally novel ones.

2. Formalism, architectures, and training objectives

Most current uMLIPs learn a scalar total energy from atomistic structure and derive mechanical observables from that learned potential-energy surface. In a universal interatomic potential for 2D materials built within the MatterSim framework, each structure is represented as a graph $$\mathcal{G} = (\mathcal{V}, \mathcal{E}, \mathcal{X}, [\mathbf{M}, \mathbf{u}])$$, with explicit three-body angular interactions implemented through spherical Bessel functions and spherical harmonics. The model predicts energy as a sum of atomic contributions,

$E = \sum_i E_i,$

with forces and stress given by

$$\mathbf{f} = -\frac{\partial E}{\partial \mathbf{x}}, \qquad \sigma = \frac{1}{V}\frac{\partial E}{\partial \varepsilon}.$$

This construction enforces translation, rotation, and permutation symmetry and is representative of the physics-aware message-passing formalism used by many uMLIPs (Wang et al., 8 Jun 2025).

A closely related decomposition appears in a universal interatomic potential for perovskite oxides, where the total energy is written as

$E_{\text{tot}} = \sum_i E_i(\text{env}_i),$

and a self-attention mechanism is used to learn the importance of neighboring atoms in chemically diverse $ABO_3$ environments (Wu et al., 2023). Across the literature, training objectives commonly combine energy, force, and stress errors, for example through weighted sums such as

$$\mathcal{L} = \lambda_E \mathcal{L}_E + \lambda_F \mathcal{L}_F + \lambda_\sigma \mathcal{L}_\sigma$$

or analogous weighted Huber or mean-squared losses (Focassio et al., 2024, Berger et al., 9 Apr 2025).

An important architectural distinction concerns whether forces are constrained to be exact derivatives of the predicted energy. In a large phonon benchmark, ORB and OMat24 were explicitly identified as models in which forces are predicted as a separate output rather than as exact energy derivatives, and this was linked to poor performance for second-derivative properties such as phonons (Loew et al., 2024). By contrast, models that compute forces by differentiation preserve energy-force consistency, which is especially consequential for vibrational properties, structural stability, and molecular dynamics.

3. Benchmark regimes and comparative performance

Phonons and vibrational thermodynamics have become one of the most discriminating UIP benchmarks. Using around 10,000 ab initio phonon calculations, one study found that MatterSim was the most accurate model for maximum phonon frequency, vibrational entropy, Helmholtz free energy, and heat capacity, with SevenNet also performing well; MACE, M3GNet, and CHGNet were described as reasonably accurate for geometry but not as strong for phonons, while ORB and OMat24 were reported to fail dramatically for phonons despite strong geometry optimization (Loew et al., 2024). A separate study built a phonon database of approximately 5,000 inorganic crystals and benchmarked 12 leading uMLIPs on Helmholtz free energy, entropy, constant-volume heat capacity, phonon frequencies, phonon density of states, and optimized atomic coordinates. In that benchmark, ORB v3 was identified as the best overall model, and the best-performing models—including ORB v3, SevenNet-MF-ompa, and GRACE-2L-OAM—were reported to achieve phonon frequency mean absolute errors below 1 meV and phonon-density-of-states Spearman coefficients above 0.94 (Han et al., 2 Jun 2025). The coexistence of these results indicates that benchmark design, reference protocol, model version, and property definition materially affect comparative rankings.

Elastic-property prediction shows a similarly differentiated landscape. A benchmark on 10,994 Materials Project structures, of which 10,871 were elastically stable at the DFT level, compared MatterSim, MACE, SevenNet, and CHGNet for elastic tensors and derived moduli. SevenNet achieved the highest overall accuracy, MACE and MatterSim balanced accuracy with computational efficiency, and CHGNet performed less effectively overall. The same study reported mean computation times per structure of 1.13 s for MACE, 1.21 s for CHGNet, 1.85 s for MatterSim, and 2.77 s for SevenNet (Gao et al., 27 Oct 2025).

Near-equilibrium equation-of-state behavior does not guarantee correct far-from-equilibrium behavior. In elemental systems, a benchmarking framework combining equation-of-state tests with minima hopping showed that no single uMLIP delivered uniformly high accuracy across all elemental groups, and that smoother learned potential-energy surfaces did not necessarily yield more accurate energetic landscapes (Tahmasbi et al., 23 Dec 2025). This result is consistent with the broader benchmark literature: UIP performance is strongly property-dependent, and accuracy for energies and relaxed geometries does not automatically transfer to phonons, elastic constants, or global structural search.

4. Applications across materials classes

The practical scope of UIPs now extends well beyond bulk crystals. For defect screening, one study used universal machine-learning interatomic potentials to perform vacancy calculations for 86,259 materials in the Materials Project database, analyze vacancy formation energies in terms of oxidation numbers, identify materials at or below the convex hull of known materials, and simulate etching of low-dimensional materials (Berger et al., 9 Apr 2025). In metals and random alloys, benchmarking on grain boundaries, general defects, high-entropy alloys, hydrogen-alloy interactions, and solute-defect interactions reported that the latest EquiformerV2 models achieved root mean square errors below 5 meV/atom for energies and 100 meV/Å for forces, outperforming specialized machine-learning potentials such as moment tensor potential and atomic cluster expansion (Shuang et al., 5 Feb 2025). For interstitial energetics in a Ti-23Nb-0.7Ta-2Zr gum metal base alloy, MACE-MATPES-PBE-0 and Orb-v3 reproduced the expected bcc site preferences—C, N, and O in octahedral sites and H in tetrahedral positions—whereas SevenNet-0 incorrectly stabilized H in octahedral coordination (Lebedaa et al., 5 Dec 2025).

Surface and interface modeling remain both important and difficult. On 1,497 surface structures from 138 bulk materials, zero-shot MACE, CHGNet, and M3GNet showed significant degradation in surface-energy prediction relative to bulk-like tasks, with systematic underestimation linked to domain shift from bulk-dominated training data (Focassio et al., 2024). By contrast, fine-tuned versions of foundation models were reported to approach the accuracy of specialized surface models, showing that universality can serve as a strong initialization even when zero-shot transfer is insufficient.

Framework materials provide a particularly stringent test because of porosity, flexible coordination, and guest-host interactions. In zeolites, all tested universal MLIPs were found to reproduce experimental or DFT-level geometries and energetics, with eSEN-30M-OAM showing the most consistent performance across pure-silica frameworks and guest-containing aluminosilicates; among analytic UIPs, GFN-FF was the best performer but did not achieve satisfactory accuracy for highly strained silica rings and aluminosilicate systems (Ito et al., 9 Sep 2025). In MOFSimBench, which evaluated more than 20 models on structural optimization, molecular-dynamics stability, bulk properties, and guest-host interactions, top-performing uMLIPs consistently outperformed classical force fields and fine-tuned machine-learning potentials across all tasks (Kraß et al., 16 Jul 2025).

Domain-specific universality has also emerged as a distinct strategy. A universal interatomic potential tailored for 2D materials was trained on 327,062 structure-energy-force-stress mappings from 20,114 materials spanning 89 elements and reported mean absolute errors of 6 meV/atom for energies, 80 meV/Å for atomic forces, and 0.067 GPa for stress tensors; it was then applied to structural relaxation, lattice dynamics, molecular dynamics, and high-throughput discovery (Wang et al., 8 Jun 2025). In perovskite oxides, a self-attention-based unified force field covering 14 metal elements reproduced experimental phase-transition sequences for several ferroelectric oxides and for the six-element ternary solid solution Pb(In$_{1/2}$Nb$_{1/2}$)O$_3$–Pb(Mg$_{1/3}$Nb$E = \sum_i E_i,$0)O$E = \sum_i E_i,$1–PbTiO$E = \sum_i E_i,$2 (Wu et al., 2023).

5. Limits of zero-shot universality

Recent work has clarified that zero-shot universality is not equivalent to quantitative reliability in all regimes. For high-temperature MOF chemistry, a benchmark dataset of 40 ps ab initio molecular-dynamics trajectories at 300, 1000, and 2000 K found that ORB-v3 and fairchem OMAT achieved the lowest energy, force, and stress errors across nine zinc- and zirconium-based MOFs, but all models exhibited significant error under high-temperature conditions. Long-timescale molecular dynamics with ORB-v3 further showed that generative error far exceeded model losses captured during static validation, with weighted loss increasing 3–4× over the static validation loss at 2000 K (Edwards et al., 28 Apr 2026).

The same concern appears in reactive and transport processes. Across seven chemically diverse systems, a comparative study found that strong performance on standard benchmarks did not guarantee accurate target observables, and that zero-shot universal MLIPs did not reliably reproduce reactive, transport, or high-barrier processes. For the sulfur-vacancy jump in MoS$E = \sum_i E_i,$3, even the best universal models underestimated the migration barrier by about 0.3 eV (Hänseroth et al., 22 Jun 2026). This result directly challenges the assumption that low global force or energy error is sufficient for quantitative mechanistic prediction.

Several broader benchmark conclusions reinforce this caution. In elemental systems, minima-hopping results revealed a decoupling between search efficiency and structural fidelity: a smoother learned potential-energy surface can improve optimization robustness while still misrepresenting the global energetic landscape (Tahmasbi et al., 23 Dec 2025). In MOFSimBench, performance across structure optimization, molecular dynamics, bulk modulus, heat capacity, and guest-host interactions was found to depend more strongly on training-set diversity and the inclusion of out-of-equilibrium conformations than on architecture alone (Kraß et al., 16 Jul 2025). A plausible implication is that “universality” is currently best understood as wide transferability under specified data and property regimes, not as unrestricted extrapolation across all chemistries, temperatures, or reaction pathways.

6. Fine-tuning, transfer learning, and emerging development paths

Fine-tuning has become a primary route for converting broad but imperfect universality into task-specific accuracy. In a study of two MACE-based foundation models, fine-tuning on task-specific datasets enhanced accuracy and in some cases outperformed models trained from scratch, while convergence was often faster because the foundation model already provided strong initial predictions. The same study emphasized dataset selection, including direct use of open-source configurations, manual filtering, and active-learning or uncertainty-based filtering (Liu et al., 9 Jun 2025). For surface modeling, conventional fine-tuning and multi-head fine-tuning both reduced prediction errors with orders-of-magnitude fewer training configurations than training from scratch, and multi-head fine-tuning preserved much more of the original bulk accuracy by combining surface and bulk objectives (Hwang et al., 30 Sep 2025).

Cross-functional transfer addresses a different bottleneck: the fidelity of the electronic-structure reference itself. Within the CHGNet framework, transfer from GGA/GGA$E = \sum_i E_i,$4 data to r$E = \sum_i E_i,$5SCAN was shown to be ineffective without energetic alignment because of large energy-scale shifts and weak raw correlations. Elemental energy referencing raised the Pearson correlation between GGA and r$E = \sum_i E_i,$6SCAN total energies from $E = \sum_i E_i,$7 to $E = \sum_i E_i,$8 and enabled more than 10-fold data efficiency relative to scratch training (Huang et al., 7 Apr 2025). A related development is direct training on higher-fidelity targets. “Matlantis-PFP v8” reported training on an r$E = \sum_i E_i,$9SCAN dataset of roughly 3 million structures and obtained improved agreement with experiments or high-accuracy references for crystals, molecules, and surfaces; in melting-point prediction, the mean absolute error was reduced from 279 K for PFP-PBE to 133 K for PFP-R2SCAN (Shinagawa et al., 9 Mar 2026).

Another emerging strategy is to use universal MLIPs as configuration-space generators rather than as final zero-shot predictors. In one-shot and iterative workflows, universal models were used to generate long molecular-dynamics trajectories, representative configurations were relabeled with DFT, and material-specific MLIPs were then trained or fine-tuned on the resulting first-principles dataset. Across the tested systems, 2,000 DFT-recalculated structures were often sufficient, and for the MoS$$\mathbf{f} = -\frac{\partial E}{\partial \mathbf{x}}, \qquad \sigma = \frac{1}{V}\frac{\partial E}{\partial \varepsilon}.$$0 sulfur-vacancy case iterative self-training recovered the DFT potential-energy profile with only 600 first-principles calculations in total (Hänseroth et al., 22 Jun 2026). At the infrastructure level, FastCHGNet demonstrated that the training cost of a universal potential can itself be compressed: the optimized implementation reduced CHGNet training time from 8.3 days on one A100 GPU to 1.53 hours on 32 GPUs, while lowering memory footprint by a factor of 3.59 and preserving model accuracy (Zhou et al., 2024).

Taken together, these results define the present state of UIP research. Universal models have become practical general-purpose tools for many atomistic tasks, especially where large-scale screening, rapid relaxation, or near-equilibrium thermodynamic and vibrational properties are required. At the same time, the most demanding regimes—surfaces, high-temperature amorphization, reactive events, and quantitatively exact target observables—still often require careful validation, explicit out-of-equilibrium training data, fine-tuning, higher-fidelity reference labels, or iterative DFT relabeling.