This article is a systems and research explainer for engineers and researchers. It is not operational or safety guidance for any specific materials workflow.

MACE vs MatterSim vs Orb: Choosing a 2026 ML Interatomic Potential

For a decade the central bargain in atomistic simulation was fixed and unhappy: you could have quantum-mechanical accuracy or you could have speed, but not both in the same calculation. Density functional theory (DFT) gave you forces good enough to trust; classical force fields gave you nanoseconds of molecular dynamics; nothing gave you both. The comparison of MACE vs MatterSim vs Orb is really a comparison of three answers to that bargain — three universal machine-learning interatomic potentials (MLIPs, models that predict energy and forces directly from atomic positions) that each claim near-DFT accuracy at a tiny fraction of DFT cost. They differ not in ambition but in the bets they place: MACE bets on symmetry, MatterSim bets on breadth of conditions, and Orb bets on raw throughput.

This is a research explainer for people evaluating these models, not deployment or safety advice for any specific simulation campaign; validate every model on your own systems before you trust a number.

What this covers: the mechanism-level differences between the three architectures, how they place on the Matbench-Discovery leaderboard, where each one quietly fails, and a practical decision procedure for picking one for stability screening, extreme-condition molecular dynamics, or large-scale high-throughput work.

Context and Background

A machine-learning interatomic potential is a regression model that maps a set of atomic coordinates and chemical species to a single scalar — the potential energy — and, by differentiating that scalar, to the forces on every atom. Train it on enough DFT-labelled snapshots and it reproduces the quantum potential energy surface (PES) to within a few tens of meV per atom while running four to six orders of magnitude faster. That is the DFT-accuracy-at-MD-cost promise, and until roughly 2022 it was delivered one system at a time: you fit a bespoke potential for silicon, or for a specific alloy, and it worked only there.

The shift that produced today’s field is the foundation MLIP — a single model pretrained on hundreds of thousands to millions of relaxation trajectories spanning most of the periodic table, meant to be used off the shelf or fine-tuned with a handful of new calculations. The Materials Project’s MPtrj dataset (about 1.6 million structures covering 89 elements, drawn from DFT relaxation trajectories) became the common training corpus; the Matbench-Discovery benchmark became the common yardstick. Once one pretrained model could relax an arbitrary inorganic crystal, a new kind of workflow became possible: closed-loop, autonomous materials discovery, where a computer proposes structures, a potential relaxes and scores them, and only the survivors ever touch a DFT queue or a robot.

That is why self-driving labs care so much about which potential they run. In a closed-loop self-driving lab architecture, the MLIP is the funnel that decides which of a million candidates deserves expensive verification. A potential that is 20 meV/atom too soft, or that mis-ranks metastable phases, silently steers the entire campaign toward dead ends. Choosing among MACE, MatterSim, and Orb is therefore not a benchmark parlour game — it is a decision about where your compute and your robots spend their time.

It helps to be precise about the magnitudes involved, because they explain why the field moved so fast. A single DFT relaxation of a modest crystal might take core-hours; a universal MLIP relaxes the same structure in milliseconds to seconds on one GPU. That is not an incremental speed-up, it is a change of kind — the difference between screening a few thousand candidates a year and screening a few million a day. The economic logic of the whole autonomous-discovery stack rests on that gap. But the gap is only useful if the fast surrogate is accurate enough that its rankings survive contact with DFT verification. If nine of every ten “stable” candidates the potential nominates turn out to be false positives, the funnel is not saving you compute, it is relocating it into a wasted verification queue. The entire value of a foundation MLIP is therefore its false-positive rate on structures it has never seen, which is exactly the quantity that varies most between MACE, MatterSim, and Orb and exactly the quantity that a single headline accuracy number hides.

MACE vs MatterSim vs Orb: the core distinction

The short answer to MACE vs MatterSim vs Orb: MACE is a higher-order equivariant message-passing network that maximises physical fidelity per parameter; MatterSim is an M3GNet-style graph network trained deliberately across a huge temperature and pressure envelope; and Orb is a deliberately non-equivariant graph network engineered for speed and scale, with newer variants that recover much of the physics the architecture nominally gives up. All three are universal potentials; they diverge on what they optimise.

To make that concrete you need one idea: equivariance. Rotate a molecule in space and its energy does not change (that is invariance), but the forces rotate with it (that is equivariance). A model is equivariant if its internal features transform correctly under rotation, reflection, and translation — the symmetry group E(3). Building that symmetry into the network by construction means the model never has to waste capacity or training data learning that physics should not depend on your choice of coordinate axes. The three models sit at different points on the spectrum from “symmetry hard-wired” to “symmetry learned from data or ignored.”

MACE: higher-order equivariance

MACE (the name stands for Multi-Atomic Cluster Expansion) is the most physics-forward of the three. It is an equivariant message-passing neural network (MPNN) — atoms are nodes in a graph, edges connect atoms within a cutoff radius, and information flows along edges in rounds of “message passing.” What makes MACE distinctive is that its messages are not just pairwise. Through the Atomic Cluster Expansion formalism it constructs higher-order messages — effectively four-body terms — in a single layer. The published consequence is striking: because each message already encodes many-body correlation, MACE needs only one or two message-passing layers where earlier equivariant models needed five or six. Fewer layers means a smaller receptive field, which means the model parallelises cleanly across a domain-decomposed simulation and stays fast despite the mathematical richness of each step. The original method paper appeared at NeurIPS 2022, and the MACE-MP-0 foundation model paper extended it to a universal potential covering 89 elements, released in variants of increasing equivariance order (L=0, 1, 2).

The pay-off of baked-in E(3)-equivariance shows up on properties that depend on subtle symmetry: phonon spectra, elastic constants, and anything sensitive to how the potential energy surface curves. Because MACE cannot represent a rotationally inconsistent force even in principle, its predictions tend to be smooth and well-behaved under the second and third derivatives that thermal and vibrational properties require. The cost is arithmetic: the tensor products that enforce equivariance at order L=2 are expensive, and MACE is generally the slowest per-atom, per-step of the three at matched accuracy.

It is worth being precise about why the higher-order messages matter, because it is the crux of MACE’s argument. Ordinary pairwise message passing describes each atom’s environment as a sum over its neighbours taken one at a time; genuine many-body correlation — the fact that three or four atoms arranged in a particular geometry have an energy that is not the sum of their pairwise pieces — has to be reconstructed by stacking many rounds of message passing. MACE instead builds those many-body terms explicitly inside a single layer through the cluster expansion, so a four-body correlation is represented directly rather than assembled from pairwise fragments. That is why it converges in one or two layers. The practical implication for a self-driving lab is subtle but real: a shallow model with a small receptive field is far easier to parallelise across many GPUs and to embed inside a large-scale molecular-dynamics engine like LAMMPS, so MACE’s mathematical sophistication does not automatically make it impractical at scale. The MACE family has also diversified — MP-0 trained on Materials Project data, MPA-0 adding the Alexandria dataset, and OMAT-trained variants — so “MACE” in 2026 is a family of checkpoints with different accuracy and license profiles rather than one model.

MatterSim: broad-condition training

MatterSim, from Microsoft Research AI for Science, makes a different bet. Architecturally the publicly released checkpoints use an internal implementation of M3GNet — a graph network with explicit three-body interactions, invariant rather than higher-order equivariant. That is a less exotic backbone than MACE’s. The distinctive move in MatterSim is the training distribution, not the architecture. The model was actively learned from large-scale first-principles calculations spanning temperatures from 0 to 5000 K and pressures up to 1000 GPa. Most foundation MLIPs are trained overwhelmingly on near-equilibrium, near-zero-Kelvin relaxation snapshots; MatterSim was deliberately fed the hot, compressed, far-from-equilibrium configurations that a real molecular-dynamics trajectory actually visits.

That design choice targets a specific weakness of the field. A potential trained only on relaxed structures has never seen a bond stretched to breaking or a lattice under 50 GPa, so it extrapolates — badly — the moment you heat it. MatterSim’s authors report near-first-principles Gibbs free energies across a wide range of inorganic solids and roughly 15 meV/atom agreement with experiment for free energies up to 1000 K, which is the quantity that actually governs phase stability at finite temperature. On its own held-out test sets the 5M-parameter model reports energy MAEs around 24-26 meV/atom on Materials-Project-like data. If your question is “what phase is stable at 900 K and 20 GPa,” MatterSim is built for exactly that regime.

Orb: fast, pragmatic architecture

Orb, from Orbital Materials, is the pragmatist. It is a graph network that is explicitly non-equivariant — it does not hard-wire rotational symmetry into its features. Classical wisdom said that was a mistake. Orb’s central empirical claim, argued at length in the Orb-v3 paper, is that the classical wisdom is wrong for practical purposes: a non-equivariant architecture, given enough data and the right regularisation, can match equivariant models on most properties while running far faster and using far less memory. Orb-v3 reports something like an order-of-magnitude latency reduction and roughly eight-fold memory reduction relative to comparably accurate equivariant models — the kind of gap that decides whether you can run a 100,000-atom simulation on one GPU.

Orb-v3 is also where the “conservative vs direct” distinction lives, and it matters enough to spell out. A conservative potential computes forces as the exact negative gradient of the predicted energy, guaranteeing energy conservation. A direct potential predicts forces with a separate output head — faster, but the forces are no longer the gradient of any energy, so long molecular-dynamics runs can slowly leak or gain energy. Orb-v2 was direct-only; Orb-v3 ships both, plus an “equigrad” trick that induces approximate rotational equivariance through a gradient-based regulariser rather than through architecture. For screening — thousands of short relaxations, throughput above all — the direct model is often the right call. For long dynamics you want the conservative variant. Orb makes that a knob you turn, and releases everything under a permissive Apache 2.0 license.

Deeper analysis: architecture, benchmarks, and cost

Under the hood all three models run the same computational pipeline; the architecture debate is about what happens in the middle box. Atoms become a graph, features propagate, a scalar energy comes out, and forces fall out by differentiation.

The mechanism worth understanding is why forces come out “for free” and why that choice has consequences. If a model predicts a single scalar energy E as a differentiable function of the atomic positions, then the force on any atom is simply the negative gradient of E with respect to that atom’s coordinates — one backward pass through the same network gives every force at once, and because they all descend from one energy, the force field is conservative by construction. This is the physically honest route, and it is what MACE always does and what Orb’s conservative variant and MatterSim do. The alternative — a direct force head — severs forces from any underlying energy. You gain speed and lose the guarantee that integrating the dynamics conserves energy, which is exactly the property long trajectories depend on. Whenever you read a benchmark, the first question is which mode produced the number.

The second mechanism is message passing itself. In each round, every atom collects “messages” from its neighbours — functions of the relative positions and the neighbours’ current features — and updates its own embedding. Stack rounds and information propagates outward: after two rounds an atom “feels” atoms two hops away. Equivariant models like MACE carry directional (tensorial) features through this process so that the messages transform correctly under rotation; non-equivariant models like Orb carry only invariant scalars and lean on data and regularisation to approximate the same behaviour. MACE’s higher-order messages let it get away with fewer rounds; Orb’s cheaper messages let it afford more atoms. Neither is universally right.

Here is where the three land on the axes that matter for materials screening. Treat the benchmark cells as directional, not gospel — see the methodology note below.

Dimension	MACE (MP-0 / MPA-0)	MatterSim (v1, 5M)	Orb (v3)
Architecture family	Higher-order equivariant MPNN (ACE messages)	M3GNet graph network (3-body, invariant)	Non-equivariant graph network
Equivariance	Built-in E(3), orders L=0/1/2	Invariant backbone	None by construction; approximate via equigrad
Training data scale	~1.6M MPtrj structures (MPA adds Alexandria)	~3M (1M model) / ~6M (5M model), wide T/P	Large multi-dataset pretraining (MPtrj, Alexandria, OMat-class)
Matbench-Discovery F1	High (MPA-0 among stronger open models; exact value — see leaderboard)	~0.83 (published)	High for conservative variants; Orb-v2 mid-pack, v3 much improved
Matbench-Discovery κ-SRME	Competitive (n/p exact here)	Competitive (n/p exact here)	Weak for direct; conservative v3 closes most of the gap
Speed / scalability	Slowest per step; scales via small receptive field	Middle	Fastest; ~10x lower latency, ~8x less memory (reported)
Energy conservation	Conservative (forces = -∇E)	Conservative	Selectable: conservative or direct
License	Code MIT; some checkpoints inherit training-data restrictions (e.g. OMat-trained variants)	MIT	Apache 2.0
Ecosystem / ASE	Mature ASE calculator, LAMMPS bindings	ASE interface, fine-tuning tooling	ASE calculator, throughput-oriented tooling

A methodology caveat is essential before anyone quotes this table. Matbench-Discovery reports several distinct metrics — F1 for stability classification (did the model correctly flag a structure as on or near the convex hull), RMSD for relaxed-geometry fidelity, and κ-SRME for how well second- and third-order force constants reproduce DFT thermal conductivity — and it now folds them into a single Combined Performance Score (CPS). These metrics do not move together. A model can top the F1 table and rank poorly on κ-SRME, because classifying stability well is a coarse task and reproducing third derivatives of the PES is a delicate one. Direct-force models are the canonical example: strong at ranking, weak at thermal transport, precisely because non-conservative forces corrupt the phonon calculation. As of mid-2026 the very top of the leaderboard is held by large equivariant transformers rather than any of our three, and the exact ordering churns month to month. The honest reading is a tier, not a ranking: MACE, MatterSim, and the conservative Orb-v3 variants are all in the strong-open-model tier; which is “best” depends entirely on which column you weight.

Cost is the axis that benchmarks under-report. On a fixed accuracy target, MACE at L=2 is typically the most expensive per atom-step, Orb the cheapest, MatterSim in between. But cost is not just wall-clock: memory sets the ceiling on system size, and here Orb’s non-equivariant design is a genuine structural advantage — no expensive tensor-product bookkeeping means you fit more atoms per GPU. If your screening funnel is dominated by many small relaxations, per-step speed dominates; if you run a few very large cells, memory dominates; the two can point at different models. This same throughput-versus-fidelity tension shows up whenever surrogate models replace expensive physics, a pattern explored in the context of neural operators for scientific simulation.

Trade-offs, gotchas, and what goes wrong

Universal potentials fail, and they fail in ways benchmarks are structurally bad at surfacing. The first and most common is out-of-distribution chemistry. Every one of these models was trained overwhelmingly on the kinds of stable, inorganic bulk crystals that populate the Materials Project. Point any of them at a chemistry the training set barely covered — organic polymers, molecular crystals, exotic interfaces, defect complexes, catalytic surfaces mid-reaction — and accuracy degrades without warning. MatterSim’s own model card is candid that accuracy for organic polymeric systems is relatively low and that surfaces and interfaces are often only qualitatively correct. The model does not announce when it is extrapolating; it just returns a confident, wrong number.

The second failure mode is over-soft potentials. Foundation MLIPs trained mostly on relaxed structures tend to under-predict the steepness of the repulsive wall — the energy penalty when atoms get too close. In practice that shows up as bonds that are too easy to compress, phase-transition pressures that come out low, and molecular dynamics that happily samples geometries DFT would forbid. A model can post an excellent formation-energy MAE and still have a systematically too-soft PES, because near-equilibrium error and high-energy repulsion are almost independent quantities.

Third: missing long-range physics. All three use a finite cutoff radius, so electrostatics and dispersion beyond that cutoff are simply absent unless bolted on. For charged defects, polar surfaces, ionic conductors, and anything where a Coulomb tail matters, a short-range MLIP can be qualitatively wrong. None of these models predicts Born effective charges or responds to an applied electric field out of the box. This is a structural limitation, not a training-data gap you can close with more examples: a message-passing network with a 5-6 Angstrom cutoff has no channel through which a charge two nanometres away can influence a force, so no amount of fine-tuning teaches it long-range electrostatics. The workarounds — appending an explicit Coulomb term, or coupling to a charge-equilibration scheme — are active research, not settled features you can assume are present.

The fourth failure mode is the non-conservative trap specific to direct-force models. Because a direct force head does not descend from any energy, the force field has a non-zero curl: you can walk a closed loop in configuration space and return with net work done, which in a long molecular-dynamics run manifests as slow, unphysical energy drift. Orb’s post-prediction corrections — mean-subtraction to zero the net force, torque elimination — patch the worst symptoms, but they do not restore true conservation. This is why direct models can look excellent on single-point benchmarks and still be the wrong tool for a nanosecond trajectory.

Fifth, and underappreciated: benchmark rank is not your-system rank. A model that wins Matbench-Discovery on the space of hull-stable inorganic crystals tells you almost nothing about its error on your specific niobate or your specific grain boundary. The reliable pattern in 2026 is fine-tuning: take a foundation model, add a few hundred to a few thousand DFT calculations from your own chemistry, and specialise it. All three support this, and it usually matters more than which base model you started from.

Practical recommendations

The honest way to settle MACE vs MatterSim vs Orb for a real project is to start from the property you actually need, not from the leaderboard, and to treat the base model as a starting point you intend to validate — never as an oracle.

For stability screening — relax a million candidates, keep the ones near the convex hull — throughput and a solid F1 win, and the direct or conservative Orb-v3 is the natural first choice, with MACE-MPA-0 as the higher-fidelity fallback when mis-ranking is expensive. For molecular dynamics at extreme temperature or pressure, or when you need finite-temperature free energies, MatterSim is the model designed for that envelope; its training distribution is the differentiator. For symmetry-sensitive properties — phonons, elastic tensors, thermal conductivity — favour MACE or a conservative Orb-v3 variant, and never trust a direct-force model for these. For very large systems where memory is the wall, Orb’s non-equivariant design is the structural advantage.

A working checklist:

Match mode to task: conservative forces for dynamics and phonons; direct forces are acceptable only for one-shot relaxations and screening.
Validate on your chemistry first: run 20-50 of your own DFT points against each candidate before committing a campaign.
Budget for fine-tuning: assume you will specialise the base model; pick one with tooling and a license that permit it.
Check the license against your use: Orb is Apache 2.0, MatterSim is MIT, MACE code is MIT but some checkpoints inherit restrictions from their training data — read the specific checkpoint’s terms before commercial use.
Log everything: capture model version, weights hash, and cutoff so results are reproducible, in line with FAIR data-provenance practice for self-driving labs, and wire the potential into your lab orchestration layer so the loop is auditable.

Frequently Asked Questions

Is MACE always more accurate than Orb because it is equivariant?

No. Equivariance guarantees the model respects rotational symmetry by construction, which helps on symmetry-sensitive properties like phonons and elastic constants. But Orb-v3’s central result is that a well-regularised non-equivariant network can match equivariant models on most benchmark metrics while running far faster. On stability-classification F1 the gap is often small; on delicate quantities like thermal conductivity the conservative, more symmetry-respecting models pull ahead. “More accurate” is only meaningful once you name the property.

What does the κ-SRME metric on Matbench-Discovery actually measure?

κ-SRME is the symmetric relative mean error in predicting lattice thermal conductivity. After a model relaxes a structure, the benchmark computes second- and third-order force constants and derives thermal conductivity, then compares to DFT. It is a demanding test because it probes third derivatives of the potential energy surface — far more sensitive than energy or force error. Direct-force models tend to score poorly here because non-conservative forces corrupt the phonon calculation, which is why the metric is reported separately from F1.

Can I use these models commercially?

It depends on the specific artifact. Orb is released under Apache 2.0, which permits commercial use. MatterSim’s released checkpoints are under the MIT License, though the model card frames the model as intended for research and experimental purposes. MACE’s code is MIT-licensed, but individual foundation checkpoints can inherit restrictions from their training data — models trained on academic-only corpora carry those corpora’s terms. Always read the license of the exact checkpoint you plan to ship.

Do I need to fine-tune, or can I use a foundation MLIP off the shelf?

For rough screening and qualitative trends, off-the-shelf is often enough. For quantitative results on a specific chemistry — especially surfaces, interfaces, defects, or anything outside the stable-inorganic-crystal training distribution — fine-tuning on a few hundred to a few thousand of your own DFT calculations usually improves accuracy far more than switching base models. All three support fine-tuning; treat it as the default for production work rather than an optional extra.

Which model handles high-temperature molecular dynamics best?

MatterSim was explicitly trained across 0-5000 K and up to 1000 GPa and reports near-first-principles free energies at finite temperature, so it is the natural first choice for hot or compressed dynamics. That said, any conservative potential validated on your system can work; the risk with the others is that their training distributions skew toward near-equilibrium structures, so they can extrapolate poorly once bonds stretch and the lattice expands. Validate at your target temperature before trusting long trajectories.

Are any of these the actual top of the Matbench-Discovery leaderboard?

As of mid-2026, no — the very top of the leaderboard is generally held by large equivariant transformer models rather than MACE, MatterSim, or Orb, and the ordering shifts as new models are submitted. All three sit in the strong-open-model tier and are widely used precisely because they balance accuracy against speed, license, and ecosystem maturity. Treat the leaderboard as a moving snapshot, not a settled verdict, and always re-check the current standings before quoting a rank.

MACE vs MatterSim vs Orb (2026): ML Interatomic Potentials