Universal MLIPs for Accurate Atomistic Modeling

• Universal MLIPs are data-driven atomistic models that approximate potential energy surfaces across diverse chemical spaces with near–first-principles accuracy.
• They employ symmetry-preserving architectures, such as equivariant message-passing networks, to encode chemical environments and capture essential physical interactions.
• They enable efficient long-time molecular dynamics and targeted fine-tuning, offering significant computational speed-ups and enhanced predictive capabilities for complex materials.

Universal machine learning interatomic potentials (MLIPs) are data-driven interatomic models intended to approximate the potential-energy surface of atoms across broad chemical spaces with near–first-principles accuracy at a fraction of the computational cost of density functional theory (DFT). In contemporary usage, “universal” denotes foundation-style models trained on chemically diverse datasets—crystals, surfaces, defects, molecules, alloys, ionic compounds, and related environments—so that a single pretrained potential can be deployed zero-shot or adapted by fine-tuning to task-specific domains. Their practical significance lies in enabling long-time, large-length-scale molecular dynamics (MD), enhanced sampling, Monte Carlo, and defect or adsorption studies that are prohibitively expensive at direct DFT accuracy, while recent work has also exposed strong domain dependence, persistent out-of-distribution failure modes, and the need for principled transfer, diagnostics, and physical augmentation (Majumdar et al., 15 May 2026, Liu et al., 9 Jun 2025, Focassio et al., 2024, Hänseroth et al., 22 Jun 2026).

1. Definition, formal structure, and symmetry constraints

At the level of mathematical form, universal MLIPs typically retain the standard atomistic decomposition of total energy into atom-centered contributions, with forces obtained by differentiation. A representative ansatz used in modern frameworks is

$$E_{\rm total}(\{r_i\})=\sum_{i=1}^N E_i\bigl(h_i(\{r\})\bigr)+E_{LR}(\{r_i\},\{q_i\}),$$

with

$\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$

Here, $h_i$ denotes a learned representation of the local environment of atom $i$, while $E_{LR}$ is an explicit long-range correction when such physics is modeled directly (Brunken et al., 21 May 2026). In MACE-style formulations the same idea appears in the simpler additive form $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$, where each $E_i$ is read from an equivariant message-passing embedding of the local environment (Liu et al., 9 Jun 2025).

The defining architectural requirement is not a particular network class but a symmetry-preserving map from atomic structure to energy and force. The common design space includes message-passing graph neural networks, equivariant tensor networks, kernel methods, and descriptor-based models. In universal settings, local representations usually encode species, neighbor distances, angles, spherical harmonics, radial basis expansions, and higher-order equivariant features. Rotational, translational, and permutational symmetries are imposed either by construction or through equivariant operations such as tensor products and Clebsch–Gordan contractions (Majumdar et al., 15 May 2026, Focassio et al., 2024).

This symmetry structure is central because universality is not merely a matter of wide elemental coverage. It also requires that the learned representation extrapolate across coordinations, phases, defects, and thermodynamic conditions. A plausible implication is that universality is best understood as a property of representation and dataset support jointly, rather than as a guarantee that a pretrained model is quantitatively reliable for every downstream observable.

2. Architectural families and software ecosystems

The dominant universal MLIP family is the equivariant message-passing model. MACE, NequIP, SevenNet, EquiformerV2, and related architectures propagate geometric information through neighbor graphs while preserving $O(3)$ or $E(3)$ symmetry. In MACE, tensor-product message passing produces high body-order features with only a small number of layers, and atomic energies are obtained by readout from the final equivariant embedding (Liu et al., 9 Jun 2025). EquiformerV2 extends this regime with an equivariant transformer backbone carrying multipole-aware features across angular momentum channels, which has been shown to deliver defect-level accuracy on broad metallic and alloy benchmarks (Shuang et al., 5 Feb 2025).

Universal MLIPs are not restricted to fully equivariant networks. CHGNet adds explicit charge and magnetic-moment inference within a graph architecture; ORB-v3 employs a scalable invariant/equivariant graph network; MatterSim combines radial and angular descriptors with attention and temperature/pressure embeddings; eSEN introduces a Mixture-of-Experts readout on top of equivariant spherical-channel features; MLANet uses a dual-path dynamic attention mechanism with multi-perspective pooling to reduce computational cost while remaining $SE(3)$-equivariant (Ito et al., 9 Sep 2025, Brunken et al., 21 May 2026, Bi et al., 24 Mar 2026).

The software layer has become increasingly important to the practical meaning of universality. The mlip v2 library was designed as a universal end-to-end framework with a single Graph data structure, Graph$\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$0Graph modules, pluggable energy heads, NVT/NPT MD, nudged elastic band, and a high-performance equivariant backend e3j; switching NequIP and MACE from pure-JAX e3nn to e3j Pallas/CUDA kernels yields up to a $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$1 reduction in end-to-end compile-and-run time (Brunken et al., 21 May 2026). AI2Pot pursues the same goal from the opposite direction: it re-engineers MTP and NEP atomistic operators as hand-crafted C++/CUDA kernels embedded in a PyTorch/Lightning workflow, reporting training speedups of $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$2 to $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$3 relative to CPU baselines and MD inference up to $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$4 k-atom-step/s at $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$5 M atoms on a single GPU (Liu et al., 8 Jul 2026).

The resulting ecosystem indicates that “universal MLIP” now refers simultaneously to a model class, a pretraining regime, and a deployment stack capable of large-scale simulation.

3. Pretraining corpora, elemental coverage, and the meaning of universality

Large-scale DFT corpora are the substrate on which universal MLIPs are built. Prominent examples include MPtrj, OMat24, Alexandria and its subsets, OFF23, OC20/OC22, OMol25, SPICE, MAD, and r$\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$6SCAN-oriented materials datasets. EqV2-based universal models have been pretrained on combinations of MPtrj, sAlex, and OMat24 totaling more than $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$7 M structures across $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$8 elements (Shuang et al., 5 Feb 2025). SevenNet-Omni is trained across fifteen databases comprising $\mathbf{F}_i=-\nabla_{r_i}E_{\rm total}.$9 M structures and thirteen DFT protocols, spanning molecules, crystals, surfaces, MOFs, and r$h_i$0SCAN channels (Kim et al., 13 Oct 2025). MACE-Osaka26 extends open universal coverage to $h_i$1 elements by combining MPtrj, OFF23, and the heavy-element HE26 dataset (Kuroda et al., 3 Mar 2026).

Elemental coverage, however, does not by itself settle the question of universality. One surface benchmark showed that openly available universal interatomic potentials—MACE, CHGNet, and M3GNet—were trained primarily on bulk materials and had no explicit surface slabs in the training data, with zero-shot surface-energy RMSEs of approximately $h_i$2, $h_i$3, and $h_i$4 J m$h_i$5, respectively, together with systematic underestimation of $h_i$6 (Focassio et al., 2024). On MOF adsorption, current universal MLIPs showed systematic over- or underbinding, and only models trained on MOF–adsorbate interactions achieved reasonable agreement with a DFT-derived reference (Edwards et al., 14 Feb 2026). In 2D high-entropy alloys, all tested universal MLIPs yielded unsatisfactory mixing energies without fine-tuning (Zhou et al., 24 Mar 2026). In reactive transport problems, strong zero-shot performance on standard benchmarks did not guarantee accurate target observables; sulfur-vacancy migration in MoS$h_i$7 remained a clear failure case (Hänseroth et al., 22 Jun 2026).

This suggests that universality in present practice is strongly conditioned by the composition of the pretraining distribution. Bulk-dominated pretraining improves broad materials transfer, but undercoordinated surfaces, reactive pathways, adsorption complexes, charged systems, and heavy elements remain partially underrepresented unless targeted data augmentation, domain bridging, or fine-tuning is introduced.

4. Fine-tuning, transfer learning, and continual adaptation

Fine-tuning is now a standard complement to pretraining. For MACE-MP-0 and MACE-MP-0b, task-specific fine-tuning improves accuracy and can outperform training from scratch, while also converging faster because the foundation model already provides strong initial predictions (Liu et al., 9 Jun 2025). Dataset selection is an explicit control variable: direct reuse of open configurations, manual filtering toward physically relevant environments, and uncertainty-based filtering via auxiliary models were all reported as viable strategies (Liu et al., 9 Jun 2025).

In chemically complex alloy design, fine-tuning has been decisive rather than optional. For $h_i$8S$h_i$9, foundation-model mixing-energy MAEs of $i$0 meV/unit and $i$1 meV/unit for two MACE variants were reduced to $i$2 meV/unit after fine-tuning on $i$3 distinct enumerated structures, while validation errors reached $i$4 meV/atom and $i$5 meV/Å (Zhou et al., 24 Mar 2026). This result directly contradicts the common misconception that a sufficiently large universal model obviates system-specific adaptation.

Continual-learning methods address a second problem: catastrophic forgetting. The reEWC framework combines Experience Replay and Elastic Weight Consolidation for fine-tuning pretrained MLIPs. On Li$i$6PS$i$7Cl, reEWC improved a pretrained model to energy MAE $i$8 meV/atom and force MAE $i$9 meV/Å with a nearly perfect slope $E_{LR}$0, while preserving performance on held-out pretraining data within approximately $E_{LR}$1 of the original model (Kim et al., 18 Jun 2025). The same study reported transfer to chemically distinct sulfide, oxide, nitride, and halide electrolytes, indicating that forgetting-aware adaptation can preserve a significant fraction of the pretrained model’s breadth (Kim et al., 18 Jun 2025).

A more radical use of universal MLIPs is to treat them as configuration-space generators rather than final predictors. In this workflow, a universal model runs long MD cheaply, the resulting frames are sub-sampled and relabeled with DFT, and a material-specific model is then trained or fine-tuned on the new labels. Across seven chemically diverse systems, $E_{LR}$2 DFT-recalculated structures were often sufficient to obtain accurate ab initio-quality models, while the most difficult MoS$E_{LR}$3 vacancy-jump case required only $E_{LR}$4 first-principles calculations in an iterative self-training loop (Hänseroth et al., 22 Jun 2026).

5. Multi-fidelity learning, electrostatics, and cross-model thermodynamics

One major limitation of universal MLIPs is their dependence on the electronic-structure level of the training labels. Multi-fidelity methods attempt to factor out this dependence. SevenNet-MF augments an equivariant GNN with fidelity one-hot embeddings, shared and fidelity-specific weights, and fidelity-dependent energy shifts/scales, allowing joint learning from PBE and r$E_{LR}$5SCAN data (Kim et al., 2024). On a universal Materials Project setting, PBE-only predictions were about $E_{LR}$6 meV/atom away from r$E_{LR}$7SCAN reference, whereas SevenNet-MF with PBE plus $E_{LR}$8 of r$E_{LR}$9SCAN data reached approximately $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$0 meV/atom under random subsampling and approximately $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$1 meV/atom under DIRECT selection; with full coverage it reached approximately $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$2 meV/atom (Kim et al., 2024).

Cross-domain transfer generalizes this idea further. SevenNet-Omni separates universal and task-specific parameters, uses selective regularization on the task parameters, and introduces a domain-bridging set sampled at about $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$3 from representative datasets to align potential-energy surfaces across tasks (Kim et al., 13 Oct 2025). In ablations, out-of-domain molecular force error at PBE level decreased from $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$4 eV/Å in a multitask model without regularization to $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$5 eV/Å with selective regularization, and to $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$6 eV/Å when combined with a multi-DBS design (Kim et al., 13 Oct 2025). The same framework reported adsorption-energy MAEs below $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$7 eV on metallic surfaces and below $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$8 eV on MOFs, while using only about $E(\{\mathbf r_i,Z_i\})=\sum_i E_i$9 r$E_i$0SCAN data overall (Kim et al., 13 Oct 2025).

A second frontier is explicit long-range electrostatics. The Latent Ewald Summation (LES) framework augments any short-range MLIP by predicting latent atom-centered charges from host-model features and adding an Ewald-summed Coulomb term so that

$E_i$1

No dipole or charge labels are required; the latent charges are inferred from energy and force supervision alone (Kim et al., 18 Jul 2025). Integrated with MACE, NequIP, CACE, and CHGNet, LES reduced force RMSE by up to $E_i$2 in bulk water short-cutoff models and restored the correct sign and magnitude of doped-versus-undoped adsorption energetics in the Au$E_i$3/MgO benchmark (Kim et al., 18 Jul 2025).

Universal MLIPs have also become objects of thermodynamic consensus building. A recent reweighting framework starts from a potential of mean force sampled under a source MLIP and analytically reweights it across target MLIPs. Because direct exponential reweighting collapses at large system size, a mean energy-gap approximation is used instead. In a $E_i$4-atom Li$E_i$5 transport problem in a nanoconfined electrolyte, this energy-only correction reproduced the target MATPES PMF structure within chemical accuracy $E_i$6 at about $E_i$7 of the full simulation cost, and reaction and activation free energies matched direct simulations within approximately $E_i$8 kJ mol$E_i$9 across PBE+D3, PBE-sol, r$O(3)$0SCAN, and r$O(3)$1SCAN-D4 levels (Majumdar et al., 15 May 2026). The same study found that MLIPs partitioned into two training-data-driven clusters, implying that consensus is not purely architectural but also functional-dependent (Majumdar et al., 15 May 2026).

6. Benchmark domains, demonstrated capability, and persistent limitations

On some benchmark classes, current universal MLIPs already approach or exceed specialized models. In metals and random alloys, EqV2-based universal models achieved $O(3)$2 meV/atom and $O(3)$3 meV/Å on nearly all defect datasets, outperforming MTP and ACE on the same BCC-W benchmarks while remaining about $O(3)$4 times faster than DFT on the reported cost metric (Shuang et al., 5 Feb 2025). In zeolites, all tested MLIPs reproduced experimental or DFT-level geometries and energetics well, with eSEN-30M-OAM giving the most consistent performance across pure silica, aluminosilicates, organic-cation, and transition-metal-containing structures (Ito et al., 9 Sep 2025).

These successes coexist with clear failure modes. Surface energies remain insufficiently accurate in zero-shot mode for predictive surface science, especially because current universal training sets are overwhelmingly bulk-dominated (Focassio et al., 2024). MOF adsorption shows systematic biases whose magnitude grows approximately linearly with CO$O(3)$5 loading, indicating compounding adsorbate–adsorbate errors (Edwards et al., 14 Feb 2026). Universal models benchmarked as configuration-space generators did not reliably reproduce reactive, transport, or high-barrier processes without subsequent relabeling and adaptation (Hänseroth et al., 22 Jun 2026). Broad element coverage also does not automatically resolve data scarcity for chemically difficult domains: heavy-element extension to $O(3)$6 elements required a dedicated HE26 dataset and changes such as larger cutoffs and spin-polarized atomic reference energies (Kuroda et al., 3 Mar 2026).

A broader critique is methodological. One position advanced in the literature is that energy and force regression on DFT trajectories is not sufficient for truly universal MLIPs. The identified gaps are overreliance on DFT as a ceiling on fidelity, insufficient metrology and interpretability beyond scalar error summaries, and the lack of computationally efficient models validated on device-scale MD for realistic materials conditions (Miret et al., 5 Feb 2025). This critique is consistent with the benchmark evidence: accuracy on held-out energies and forces is necessary but not sufficient for reliable prediction of barriers, adsorption thermodynamics, phase stability, or long-time MD stability.

7. Outlook

Current research points toward an overview of broader pretraining, sharper physical inductive bias, and more disciplined transfer. Cross-domain multi-head replay training has already shown that a single unified MACE-derived model can maintain state-of-the-art materials accuracy while substantially improving molecular and surface performance through shared backbone learning (Batatia et al., 29 Oct 2025). Multi-fidelity learning suggests that relatively small high-fidelity fractions can steer a universal model toward r$O(3)$7SCAN- or even coupled-cluster-level targets when low-fidelity coverage is large (Kim et al., 2024). Electrostatic plug-ins such as LES indicate that some long-range physics can be recovered without redesigning the host architecture (Kim et al., 18 Jul 2025).

Several directions therefore appear structurally important. One is dataset diversification toward surfaces, interfaces, defects, liquids, adsorption complexes, and charged or magnetic systems, rather than further scaling of bulk-heavy corpora alone (Focassio et al., 2024, Batatia et al., 29 Oct 2025). Another is continual-learning machinery that preserves pretrained breadth during domain specialization (Kim et al., 18 Jun 2025). A third is model metrology: uncertainty proxies, out-of-domain diagnostics, free-energy reweighting, and direct benchmarking on derived observables rather than only on instantaneous labels (Majumdar et al., 15 May 2026, Miret et al., 5 Feb 2025). Finally, hardware-aware software and specialized kernels remain essential if universal MLIPs are to approach the throughput needed for million-atom, nanosecond-scale simulation campaigns (Brunken et al., 21 May 2026, Liu et al., 8 Jul 2026).

Taken together, the field has moved from narrowly trained interatomic surrogates toward genuine foundation-style atomistic models. Yet the available evidence also shows that present universality is conditional: it is strongest where data coverage, symmetry-aware architecture, physical augmentation, and transfer protocol are jointly aligned, and weakest where any one of those ingredients is missing.