Universal MLIPs: Pretrained Atomistic Models

• Universal MLIPs are foundational models that map atomic environments to energies, forces, and stresses across diverse chemistries.
• They leverage graph-based, symmetry-aware architectures and extensive training datasets to enable zero-shot predictions and transferable learning.
• Adaptation strategies like fine-tuning, transfer learning, and knowledge distillation address limitations in reactive or extreme regimes.

Universal machine-learned interatomic potentials (u-MLIPs) are pretrained, foundation-style atomistic models that map atomic positions and species to total energies, forces, and often stresses across many elements, bonding types, and phases without system-specific retraining. In the standard formulation, the total energy is expressed as $E_{\text{tot}} = \sum_i E_i(\mathcal{N}_i)$, where $\mathcal{N}_i$ denotes the local environment of atom $i$, and forces follow from $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$. Contemporary u-MLIPs are trained on very large, chemically diverse datasets and are intended to function both as zero-shot interatomic models and as reusable priors for fine-tuning, transfer learning, active learning, and configuration-space generation (Hänseroth et al., 22 Jun 2026, Focassio et al., 2024, Shuang et al., 5 Feb 2025).

1. Definition and conceptual scope

The defining distinction between u-MLIPs and system-specific MLIPs is not architecture but training regime and deployment regime. Universal models are trained once across many elements and chemistries and are meant to be applied off-the-shelf to new systems; material-specific MLIPs are trained or fine-tuned on small, system-specific datasets and are targeted to reproduce particular observables such as diffusion barriers, ionic conductivities, proton transport, or defect energetics (Hänseroth et al., 22 Jun 2026). This distinction has become increasingly important because the same architecture family—MACE, SevenNet, GRACE, MatterSim, ORB, PET, CHGNet, M3GNet, or PFP—can appear either as a foundation model or as a local, system-adapted model.

Within the recent literature, “universality” is operational rather than absolute. It usually denotes broad elemental coverage, broad configurational coverage, and intended transfer across many crystal structures and chemistries without retraining, but it does not imply uniformly reliable zero-shot accuracy in every thermodynamic regime or for every target observable (Focassio et al., 2024). Several studies therefore treat u-MLIPs as foundations rather than final production force fields: they are strong priors, efficient samplers, and useful initialization points, but they remain bounded by both training-set coverage and the fidelity of the reference electronic-structure method (Hänseroth et al., 22 Jun 2026, Yu et al., 2024).

A recurrent conceptual shift in the field is the move from viewing universality as “one model replaces all specialized models” to viewing it as a hybrid capability. In this view, u-MLIPs accelerate atomistic workflows by supplying broad chemical priors, cheap long-time molecular dynamics, and transferable parameter initialization, while quantitative predictions for demanding observables are recovered through relabeling, fine-tuning, or higher-fidelity transfer (Hänseroth et al., 22 Jun 2026).

2. Architectural families and training corpora

Most current u-MLIPs are graph-based and symmetry-aware. The dominant architectural classes include equivariant message-passing GNNs such as MACE, SevenNet, PET, and many ORB variants; graph transformers such as EquiformerV2 and eSEN; graph-network simulators such as orb; graph-basis atomic cluster expansion models such as GRACE; and multi-mode universal potentials such as PFP v8 built on TeaNet, a tensor-equivariant message-passing GNN carrying scalars, vectors, and rank-2 tensors (Oh et al., 13 Apr 2026, Shinagawa et al., 9 Mar 2026).

Their training corpora are correspondingly broad. Materials Project trajectories, Alexandria, OpenMaterials2024, OpenMolecules2025, OMAT24, ODAC23, OC20/22, MatPES, MP-r$^2$SCAN, and large proprietary corpora recur as the dominant data sources, with differing emphases on bulk crystals, molecules, adsorbates, surfaces, defects, disordered structures, and high-temperature snapshots (Hänseroth et al., 22 Jun 2026, Oh et al., 13 Apr 2026). A central empirical theme is that the breadth of the dataset often matters as much as, or more than, architectural novelty.

Model family	Representative models	Characteristic training scope
Equivariant message-passing GNNs	MACE, SevenNet, PET, MatterSim	MPtrj, OMAT24, Alexandria, MatPES
Graph transformers / equivariant large models	EquiformerV2, eSEN	OMat24, MPtrj, sAlex
Graph-network simulators / graph-basis ACE	orb, GRACE	MPtrj, Alexandria, OMat24
Multi-mode universal potentials	PFP v8	PBE, PBE+U, r$^2$SCAN, $\omega$B97X-D, OC20 modes

The training target itself is now a major design axis. Most first-generation universals emulate PBE or PBE+U. More recent work has shown that cross-functional transfer is nontrivial because raw GGA/GGA+U and r$^2$SCAN total energies are poorly correlated: in CHGNet, the Pearson correlation is $\rho = 0.0917$ for raw total energies, but rises to $\rho = 0.9250$ after elemental energy referencing with AtomRef (Huang et al., 7 Apr 2025). PFP v8 addresses the same issue from the other direction by training directly on r$\mathcal{N}_i$0SCAN across crystals, molecules, slabs, and disordered structures, with calculation-mode conditioning and optional D3 correction handled analytically at inference (Shinagawa et al., 9 Mar 2026). This suggests that universality is increasingly understood as simultaneous coverage of chemical space, configurational space, and fidelity space.

3. Benchmark evidence: strengths and failure modes

The strongest consistent result across benchmarks is that zero-shot u-MLIPs are often highly competent near equilibrium but substantially less reliable for out-of-domain, reactive, or barrier-dominated phenomena. The most explicit formulation is that “good benchmark metrics ≠ good target observables” (Hänseroth et al., 22 Jun 2026). In MoS$\mathcal{N}_i$1 with sulfur vacancies, older foundation models fail even qualitatively for the vacancy-migration barrier, and the best zero-shot universal model in that study, PET-OAM-XL, underestimates the barrier by $\mathcal{N}_i$2 eV; the same work shows that energy errors below $\mathcal{N}_i$3 eV/atom and acceptable global force errors still do not guarantee correct reactive dynamics (Hänseroth et al., 22 Jun 2026).

Surface science reveals an analogous pattern. On a six-element subset, zero-shot MACE, CHGNet, and M3GNet yield surface-energy RMSEs of $\mathcal{N}_i$4, $\mathcal{N}_i$5, and $\mathcal{N}_i$6, respectively, and all three typically underestimate surface energies, a systematic error described as PES “softening” (Focassio et al., 2024). In high-temperature MOF chemistry, fairchem OMAT and ORB-v3 give the best static errors among five universal models, with fairchem OMAT reaching an overall $\mathcal{N}_i$7 meV/atom energy MAE and $\mathcal{N}_i$8 meV Å$\mathcal{N}_i$9 force MAE, yet all models degrade strongly at $i$0 K, and the generative error in long MD can be $i$1–$i$2 larger than static validation loss (Edwards et al., 28 Apr 2026). Under pressure, performance also degrades systematically: for example, M3GNet’s energy MAE grows from $i$3 to $i$4 meV/atom between $i$5 and $i$6 GPa, and volume and energy errors for most models worsen as compressed environments move outside the ambient-pressure training distribution (Loew et al., 25 Aug 2025).

Global PES exploration exposes a related but distinct limitation. In elemental Minima Hopping benchmarks, EOS behavior near equilibrium and far-from-equilibrium structure-search behavior decouple: smoother learned PESs do not necessarily yield more accurate energetic landscapes, and low instability in local optimization does not guarantee recovery of the correct set of minima or the correct relative energy ordering (Tahmasbi et al., 23 Dec 2025). In nanoporous materials, MOFSimBench arrives at a parallel conclusion: off-the-shelf universal models can already outperform UFF, UFF4MOF, and several fine-tuned baselines on structural optimization, MD stability, bulk modulus, heat capacity, and guest–host energetics, but non-conservative models can remain unstable and broad performance still depends strongly on the training distribution (Kraß et al., 16 Jul 2025).

4. Adaptation strategies: fine-tuning, transfer, and distillation

The leading practical response to these failure modes is not abandonment of u-MLIPs but systematic adaptation. One route is to use universal models as configuration-space generators. A representative workflow runs $i$7 ns of inexpensive MD with a state-of-the-art u-MLIP, equidistantly subsamples $i$8 frames, relabels all frames with DFT single points, and then trains or fine-tunes a material-specific MLIP on those labels. Across seven chemically diverse systems, $i$9 DFT structures are often sufficient to obtain accurate material-specific models, and for the hardest MoS$$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$0 case an iterative self-training loop recovers the DFT barrier profile with only $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$1 first-principles calculations in total (Hänseroth et al., 22 Jun 2026).

A second route is direct domain adaptation. For surfaces, fine-tuning MACE-MP-0 on a $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$2-structure surface-rich dataset reduces surface-energy RMSE from $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$3 to $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$4, demonstrating that relatively small, task-specific datasets can substantially repair systematic softening (Focassio et al., 2024). More broadly, a dedicated fine-tuning study of MACE-MP-0 and MACE-MP-0b finds that fine-tuning often converges faster than training from scratch and can outperform scratch models, but that the outcome depends strongly on dataset selection, filtering, or active-learning-style curation (Liu et al., 9 Jun 2025).

Cross-functional transfer has its own specialized recipe. Within CHGNet, naïve GGA$$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$5r$$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$6SCAN transfer is ineffective, but replacing the source AtomRef with an r$$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$7SCAN AtomRef before fine-tuning lowers test energy MAE from $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$8 to $$\mathbf{F}_i = -\partial E_{\text{tot}}/\partial \mathbf{R}_i$$9 meV/atom, force MAE from $^2$0 to $^2$1 meV/Å, and stress MAE from $^2$2 to $^2$3 GPa, while also improving decomposition and formation energies (Huang et al., 7 Apr 2025). This establishes elemental energy referencing as a core ingredient of multi-fidelity universal MLIPs.

A third route is knowledge distillation. SevenNet-Nano distills a $^2$4M-parameter SevenNet-Omni teacher into a $^2$5k-parameter student, yielding a model that is $^2$6 smaller and attains up to $^2$7 speedup at large system size while preserving broad transferability across crystals, defects, molecules, MOFs, and extreme nonequilibrium plasma-etching trajectories (Oh et al., 13 Apr 2026). The same study shows that task-specific fine-tuning of the distilled student can be highly data-efficient, but also that replay is required to avoid catastrophic forgetting in extreme regimes.

5. Established application domains

The application portfolio of u-MLIPs is now broad enough that no single benchmark captures it. In crystal structure prediction, M3GNet-driven USPEX searches rediscover known quaternary oxides absent from training and identify seven new thermodynamically and dynamically stable or metastable compounds in Sr–Li–Al–O and Ba–Y–Al–O; the same study reports a $^2$8-fold speedup for a representative structure-prediction task, while also showing that higher-level validation with SCAN, R2SCAN, or RPA remains essential because a PBE-trained u-MLIP inherits PBE’s phase-ordering biases (An et al., 3 Feb 2026).

In metallic defects and random alloys, the best EquiformerV2 models achieve root mean square errors below $^2$9 meV/atom for energies and $^2$0 meV/Å for forces on comprehensive defect datasets, including grain boundaries, dislocations, high-entropy alloys, hydrogen–alloy interactions, and solute–defect interactions, outperforming specialized moment tensor and atomic cluster expansion potentials in several cases (Shuang et al., 5 Feb 2025). This is one of the clearest demonstrations that current u-MLIPs can act as practical DFT surrogates in a major application class.

For vibrational properties, a phonon benchmark over nearly $^2$1 inorganic crystals shows that ORB v3 reaches a mean phonon-frequency MAE of $^2$2 meV, a mean phonon-DOS Spearman coefficient of $^2$3, and a mean free-energy MAE of $^2$4 meV/atom, enabling fast phonon calculations and real-time or near-real-time interpretation of inelastic neutron scattering data (Han et al., 2 Jun 2025). In MOF adsorption screening, a hybrid UFF+PFP workflow identifies cases where a u-MLIP is essential to capture hydrogen bonding, confinement, and flexibility: for $^2$5 candidate MOFs, PFP matches DFT interaction energies with MADs of $^2$6 kJ mol$^2$7 for ethylene and $^2$8 kJ mol$^2$9 for water, while full cell relaxation can shift ethylene affinity by up to $\omega$0 kJ mol$\omega$1 in flexible frameworks (Bonakala et al., 8 Sep 2025).

Extreme nonequilibrium applications are also beginning to enter the universal domain. SevenNet-Nano reproduces plasma-etching yields in SiO$\omega$2 after teacher-guided fine-tuning and preserves short-range repulsion in quasi-static drag tests that reach the $\omega$3–$\omega$4 eV regime (Oh et al., 13 Apr 2026). PFP v8 extends the same trend to long-time thermodynamics by predicting melting points with an average error of approximately $\omega$5 K, roughly halving the error relative to PBE-trained models while improving zero-shot agreement with experiment across crystals, molecules, and surfaces (Shinagawa et al., 9 Mar 2026).

6. Limitations, controversies, and research directions

The main controversy surrounding u-MLIPs is not whether they are useful, but what “universal” should mean in practice. One position treats strong zero-shot performance as the end goal; another, increasingly supported by benchmarks, treats u-MLIPs as reusable but imperfect priors whose most robust role is to accelerate sampling, initialization, and hierarchical modeling. The latter view is reinforced by several independent observations: zero-shot models mis-handle surface energies, high-temperature MOF chemistry, high-pressure compression, and some reactive barriers; functional biases of PBE-trained models propagate directly into downstream predictions; and even excellent static metrics can conceal large generative errors in long MD (Edwards et al., 28 Apr 2026, Focassio et al., 2024).

A second controversy concerns whether current limitations are primarily architectural or data-driven. Multiple studies explicitly argue for the latter. Surface benchmarks attribute the dominant gap to bulk-centric training data rather than to lack of model capacity (Focassio et al., 2024). High-pressure and high-temperature MOF benchmarks make the same point for compressed environments, bond breaking, amorphization, and pyrolysis (Loew et al., 25 Aug 2025, Edwards et al., 28 Apr 2026). MOFSimBench generalizes the argument further by showing that the decisive improvement across generations of universal models tracks the inclusion of out-of-equilibrium configurations such as OMat24 more clearly than any single architectural shift (Kraß et al., 16 Jul 2025). This suggests that a “truly universal” training corpus would need systematic coverage of surfaces, interfaces, defects, amorphous phases, extreme pressures, high-temperature reactive states, and multicomponent disorder.

A third issue is calibration beyond PBE. CSP results on Sr$\omega$6LiAlO$\omega$7 show that a PBE-trained universal potential can correctly reproduce the PBE ground state while still being wrong with respect to SCAN, R2SCAN, and RPA (An et al., 3 Feb 2026). PFP v8 addresses this by promoting r$\omega$8SCAN to a primary training target and explicitly making experimental agreement a design objective (Shinagawa et al., 9 Mar 2026). This direction, together with AtomRef-based cross-functional transfer, points toward multi-fidelity and multi-functional foundation potentials rather than a single monolithic PBE surrogate.

Finally, reliability assessment remains incomplete. Ensemble uncertainty estimation in large defect benchmarks provides some information for energies but underestimates force errors (Shuang et al., 5 Feb 2025). High-temperature MOF work argues that generative testing should become standard, because static validation can substantially understate trajectory error (Edwards et al., 28 Apr 2026). The likely near-term outcome is therefore a layered practice: universal zero-shot inference for screening and initialization, targeted DFT validation on task-relevant observables, and lightweight adaptation—fine-tuning, relabeling, replay-based distillation, or active learning—when the target regime sits outside the statistical center of the foundation dataset.