MACE Machine-Learned Potentials: From Equivariant Message Passing to Atomistic Foundation Models

New
License: MIT (code); some weights are released under the Academic Software License (ASL) for academic use only · Updated 2026-11-12
My rating

1. Overview

Machine-learned interatomic potentials have changed what atomistic simulation can afford, yet two constraints kept them out of routine engineering use for years: the human cost of developing and validating a potential for each chemical system, and the poor transferability of potentials trained on one chemistry when applied to the next. MACE addresses both by training once, over broad coverage of elements and chemistry, and letting a user adapt the result with a small amount of application-specific data. The repository is the open-source project of the Csanyi group at the University of Cambridge, hosted at github.com/ACEsuit/mace. Its state in November 2026: created on 21 June 2022, last pushed on 11 November 2026, 1411 stars, 514 forks, 278 open issues, default branch develop, primary language Python, roughly 190 MB in size, not archived.

Three reasons justify a deep entry. First, MACE is among the most cited implementations of the equivariant machine-learning potential line of work. The NeurIPS 2022 paper titled MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields, with DOI 10.52202/068431-0830 and five authors led by Ilyes Batatia, established the architecture, and the same group's foundation-model work, A foundation model for atomistic materials chemistry, with arXiv identifier 2401.00096 submitted on 29 December 2023 with 88 authors and a journal version in The Journal of Chemical Physics under DOI 10.1063/5.0297006 dated 13 November 2025, extended it into a general-purpose model. Second, the code organisation is directly instructive for teams building physical-AI tooling: the model registry and weight download logic live in first-class modules, so introducing a new generation of foundation models does not require changes on the calling side. Third, the project has grown a dual-ecosystem structure, with a Python side covering the main training and simulation workflow and a Julia side providing a high-performance alternative, together forming 63 public repositories.

A caution is warranted. MACE-MP-0 is only the first generation of weights in this family. Later releases added MACE-MP-0b, 0b2 and 0b3, MACE-MPA-0, MACE-OMAT-0, the MACE-MATPES family, MACE-MH-0 and MACE-MH-1, MACE-OMOL-0 and MACE-POLAR-1. Treating the entry as a single model understates its coverage and does not explain its sustained adoption.

2. How It Works

The MACE architecture merges atomic cluster expansion with equivariant message passing inside a single computation graph. Equivariance means that the model output transforms in a well-defined way when the input system is rotated: scalar quantities such as energy are invariant, vector quantities such as forces rotate with the coordinates, and equivariant channels transform according to their angular momentum order. Equivariant potentials converge faster from small datasets, because samples that differ only by rotation need not be treated as independent inputs.

The distinguishing feature is body order. The original cluster expansion constructs basis functions explicitly up to four-body interactions, and the parameter count grows combinatorially with body order, which limited practical use in production molecular dynamics. MACE instead uses high-order equivariant features at every message-passing layer, encoding high-order many-body information into feature dimensionality rather than into expansion order, which brings the number of interaction layers down to two. The NeurIPS 2022 abstract describes this as reaching higher accuracy with fewer interaction layers, later demonstrated on benchmarks including rMD17 and 3BPA.

The computation proceeds in five steps. A graph is built over each atom's neighbours within a cutoff radius, with radial distances on the edges. Initial features are projected by mapping scalar attributes such as atomic number, group and period through a linear layer into a per-channel radial expansion that yields scalar, vector and higher-order equivariant channels. Equivariant message passing then runs: neighbours generate messages through a radial multilayer perceptron whose angular momentum orders match the target channels, so rotational covariance is preserved; aggregated messages update node features through a linear combination. Readout contracts each atom's channel features to a scalar energy contribution and sums them to obtain the total system energy. Forces follow from automatic differentiation of that energy, so energy and force training are two outputs of a single objective and the conservation laws are satisfied by construction. An energy shift then corrects the isolated-atom reference, which is the theoretical zero of the potential and the source of a systematic bias at low density if left untreated.

Training and use are organised around the Atomic Simulation Environment. The core calculator is a standard ase.calculators.calculator implementation, so it can be assigned to atoms.calc and connected to any ASE-based molecular dynamics, geometry optimisation or thermodynamics workflow. The calculators directory contains the core implementation, a LAMMPS interface for the libtorch format, a second LAMMPS interface built on the MLIAP interface introduced in v0.3.13, and a torch-sim integration module. This organisation means the marginal cost of connecting a MACE potential to an existing simulation pipeline is low, and physical-AI teams typically need only a few dozen lines to replace a commercial solver or a bespoke DFT path.

3. Scope and Assumptions

The scope of MACE can be drawn along three axes: system type, data density and physical-quantity coverage.

On system type, the MACE-MP series was trained on the MPtrj trajectory dataset from the Materials Project, covering 89 chemical elements in inorganic crystal configurations under the PBE+U functional, which makes it a natural fit for lattice dynamics, phonon spectra, phase stability and mechanical properties of inorganic solids. For molecular systems, MACE-OFF23 is the version trained specifically for small organic molecules, covering 83 elements and built on a large organic dataset, aimed at binding affinity, conformational sampling and free-energy perturbation. Surfaces and catalysis are supported through the OC20 dataset, and the MACE-MH series uses a multi-head architecture to cover bulk, molecular and surface targets simultaneously.

Data density is the second constraint. The central selling point is that a small amount of target-system data suffices, since fine-tuning on a handful of configurations approaches first-principles accuracy. That promise carries a condition: the fine-tuning data must come from the same electronic-structure method as the target, because PBE and r2SCAN are not interchangeable and a mismatch leaves a systematic method error after fine-tuning. The MACE-MH-1 release addresses this by providing six parallel heads spanning PBE, r2SCAN, wB97M-V and B3LYP targets, turning cross-theory transfer into an architectural feature rather than a manual model swap.

Physical-quantity coverage is the third boundary and the one most easily overlooked. Most MACE weights predict energy, force and, in some versions, stress, but not charge, spin, dipole moment or polarizability. MACE-MDP, published as a ChemRxiv preprint under DOI 10.26434/chemrxiv.15000716/v1 on 9 March 2026, targets dipole moments and polarizabilities only and the repository states plainly that it must not be used for energies and forces. MACE-POLAR-1, introduced in v0.3.16 in May 2026, adds explicit long-range electrostatics through a non-self-consistent polarisable field, learning per-atom charge and spin densities from energy and force labels alone, but it covers 83 elements and one specific dataset.

Licensing forms a fourth boundary that carries real consequence. The code itself is standard MIT, with LICENSE.md carrying the MIT License and a 2022 copyright line, but weight licences are not uniform. The model table in the foundations repository lists MACE-MP-0a, 0b, 0b2, 0b3 and MACE-MPA-0 as MIT, while MACE-OMAT-0, the MACE-MATPES models, MACE-MH-0 and MACE-MH-1, MACE-OMOL-0 and MACE-POLAR-1 are under the Academic Software License. The MACE-OFF23 repository states it directly: downloading the models constitutes agreement to the Academic Software License, free for academic purposes but not for commercial purposes. In a commercial engineering delivery context the code can be used freely while a substantial share of the newest weights require separate clearance. This is the easiest link in the chain to overlook during tool selection.

4. Getting Started

There are two installation routes. The PyPI distribution name is mace-torch, not mace. The package called mace on PyPI was uploaded in 2017, has four releases, and is unrelated to this project, which makes it the most common install mistake. As of 27 October 2026 the latest mace-torch version is 0.3.17, with 15 releases in total, requiring Python 3.11 or newer and PyTorch 2.5 or newer, and depending on e3nn 0.4.4, numpy, opt_einsum, ase, torch-ema, prettytable, matscipy, h5py and torchmetrics.

The minimal working example loads a calculator through ASE:

from ase.build import bulk
from mace.calculators import mace_mp

atoms = bulk("Cu", "fcc", a=3.6, cubic=True)
atoms.calc = mace_mp(model="medium", device="cpu")
print(atoms.get_potential_energy(), atoms.get_forces())

Model size is selected through the model parameter, with small, medium, large and extra-large options. Weights are downloaded on first use and cached, using normalised raw.githubusercontent.com URLs and a 120 second socket read timeout.

Fine-tuning is the second step, used to adapt a general weight to a concrete system. The pattern is to sample configurations of the target system with a general model first, ensuring the fine-tuning set covers the relevant distribution, then train from a foundation model checkpoint with the label columns specified explicitly. Pseudo-label fine-tuning added in v0.3.14 allows a model to continue training on its own predictions, which stabilises convergence when high-accuracy labels are scarce.

The third step is production-scale molecular dynamics. Two LAMMPS paths are provided, one using the libtorch format and one using the MLIAP interface introduced in v0.3.13 with CuEquivariance GPU acceleration, selected through the format flag of the model conversion command. v0.3.13 also added multi-GPU inference through MPI message passing of intermediate tensors, plus atomic stress and atomic virial output.

Version alignment is the last engineering step. The v0.3.17 release notes explicitly warn about options that v1.0 removes or replaces, including a calculator warmup parameter that was accepted but did nothing and now raises a TypeError, IPEX extension support, a CUDA 11 extra that no longer matches the frontend, and the training-time default that disables convolution fusion after it regressed badly on variable-shape batches while inference keeps it enabled.

5. Comparison with Alternatives

Placing MACE in the landscape of machine-learned potentials requires three comparison axes: architectural line, weight strategy and ecosystem shape.

On architecture, equivariant message passing became the dominant consensus after 2022. Direct competitors include NequIP, Allegro, SevenNet, CHGNet and its successor model families, MatGL and the Deep Potential stack. The differentiator for MACE is the high-order equivariant feature set that allows a two-layer interaction structure, which converts directly into inference speed, and interaction depth is what determines per-step cost in long production trajectories.

Weight strategy is the second axis and the one where MACE is genuinely distinctive. Many machine-learning potential projects publish training code and scripts and leave users to train their own weights. MACE established a dedicated foundations repository in January 2024 that distributes pretrained weights, with ten releases spanning the model generations from the first MP weights through the later multi-head, molecule, hybrid-functional and polarisable entries. The same weights are mirrored on Hugging Face. Publishing code, weights, training scripts and dataset citations together is the standard shape of a foundation model release.

Ecosystem shape is the third axis. MACE is one of the few machine-learning potential projects maintaining parallel Python and Julia implementations, where the Python side handles mainstream simulation integration and training and the Julia side offers higher performance together with differentiable Hamiltonian construction across several dozen packages. The organisation also hosts a JAX backend and, more recently, a second JAX project, indicating an ongoing migration of equivariant architectures toward compiled automatic differentiation. Tutorials, a model registry and a potential registry complete the tooling.

One point deserves objective statement. Several Julia packages carry no explicit licence or a non-standard one in the GitHub metadata, in contrast to the clearly licensed Python main repository, so commercial adoption requires checking each package individually.

6. Benchmarks and Evidence

The headline evidence for MACE-MP-0 comes from the foundation-model paper. Its abstract states the central result: a general-purpose atomistic model trained on a public dataset of moderate size can run stable molecular dynamics across a wide range of molecules and materials, and the paper demonstrates the capability of MACE-MP-0 on the dynamics of solids, liquids, gases, chemical reactions, interfaces and a small protein, with qualitative and at times quantitative accuracy.

The benchmarks fall into three families. The first is the small-molecule set of rMD17 and 3BPA, used as architectural validation. The second is the materials-stability framework Matbench Discovery, arXiv 2308.14920 from 28 August 2023 with ten authors including Gerbrand Ceder, which applies machine-learned energy models as pre-filters for high-throughput searches of stable inorganic crystals and maintains a Python package and an online leaderboard. The foundations repository records that MACE-MPA-0 achieves state-of-the-art accuracy on the Matbench benchmarks and improves substantially over the earlier MACE-MP-0 models on material systems. The third family is the joint comparison of energy, force and stress errors, where the MACE-MATPES release notes tabulate MACE against M3GNet, CHGNet and TensorNet under the PBE functional.

On the data itself, MPtrj derives from Materials Project trajectories computed with the DFT PBE+U functional, comprising on the order of 150 thousand inorganic crystal configurations covering most of the periodic table. The training archive is downloadable from the foundations repository, and users are asked to cite the CHGNet paper as the dataset origin. MACE-MPA-0 was trained on the combination of MPtraj and sAlex, adding three and a half million crystal configurations.

The MACE-OFF23 data chain runs separately, with training and evaluation data deposited in the Cambridge Data Repository under DOI 10.17863/cam.107498, a 2024 dataset record, and the corresponding article is MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules in the Journal of the American Chemical Society under DOI 10.1021/jacs.4c07099 dated 19 May 2025, with preprint version arXiv 2312.15211.

For third-party perspective, a Nature Reviews Chemistry article, Foundation models for atomistic simulation of chemistry and materials, DOI 10.1038/s41570-025-00793-5, published 11 February 2026 in volume 10 on pages 212 to 230 with 14 authors led by Eric C.-Y. Yuan, surveys the whole field of atomistic foundation models and is a high-value reference for positioning MACE. Chemical Science in 2026 published Node-equivariant message passing for efficient and accurate machine learning interatomic potentials, DOI 10.1039/d5sc07248d, volume 17, pages 3793 to 3803, representing a direction of architectural improvement after MACE that generalises per-atom equivariance to node-level equivariance.

7. Role in Physics-AI Workflows

The role MACE plays in a physical-AI workflow can be stated in one line: it pushes first-principles accuracy down to a time scale that engineering can afford, which is what makes large volumes of high-quality physical data producible in the first place. It is the coupling point between data-driven methods and physics.

The first application is as a label source for surrogate models. Training downstream models for material property prediction, potential energy surface fitting, ensemble dynamics or rare-event sampling is dominated by label cost, and the stability of the training examples sets the ceiling on the surrogate. Labels generated with MACE reach the order of hundreds of thousands of configurations spanning crystals, liquids and surfaces. This three-stage loop of machine-learning potential generated labels, surrogate learning and surrogate-guided sampling back into Monte Carlo or molecular dynamics has become standard practice, and it connects to the site's clusters on data assimilation and Bayesian inversion and on surrogate modelling and reduced-order models.

The second application is the design and control loop. For morphology optimisation, interface engineering or catalytic site screening, the objective function must be evaluated across many candidate configurations, and DFT cost per configuration makes enumeration infeasible. Lowering evaluation cost to the molecular-dynamics scale makes a loop of generate, evaluate, screen and refine with DFT practical. This path has direct engineering meaning for material selection in power and computing hardware, including electrode materials and thermal interface materials, and it pairs naturally with the topology optimisation entry in this site.

The third application is use as an environment model inside reinforcement learning or active learning cycles, where contact mechanics and fracture in robotic material handling are hard to describe analytically and a machine-learned potential serving as a high-fidelity environment accelerates policy search. This path places a high demand on extrapolation safety, since configurations outside the training distribution must be detected and rejected.

The fourth application is the intermediate layer in cross-scale coupling. MACE supplies the atomistic scale while engineering problems span atomistic, mesoscopic and macroscopic levels, so atomistic outputs such as force and stress can feed mesoscale finite element or phase-field models, replacing constitutive relations and fracture criteria with MLIP-derived fields and providing more realistic material parameters at the macroscopic level. This direction contrasts with the continuum modelling discussed in the bridge structural simulation entry.

Selection should follow a fixed order: check first whether the target system lies inside the training distribution of some weight, across element coverage, functional and structure type; then confirm the required physical quantities are within the weight output; then check whether the weight licence suits the use case; and only then assess whether fine-tuning is needed and whether high-accuracy labels are available. If any of the first three checks fails, going directly to fine-tuning usually works better than forcing a pretrained weight onto the problem.

8. Limitations and Failure Modes

The failure modes have identifiable technical causes, and each maps to a practical check.

Out-of-distribution configurations cause divergence. A machine-learned potential has no physical constraint outside its training distribution and may return non-monotonic energies or exploding forces. Typical triggers include element combinations absent from the training set, extreme compression or high temperature producing unusual coordination, transition-state configurations during bond breaking and forming, and systems whose charge or spin state changes sharply. The engineering defence is a sanity check on energy and force at every molecular dynamics step, aborting rather than continuing when non-physical values appear.

Functional mismatch gives systematic method error. Pretrained weights are bound to a specific electronic-structure method, so PBE weights cannot substitute for r2SCAN weights in a high-accuracy calculation, and using them where r2SCAN accuracy is required yields a systematic method error rather than a slightly worse but comparable result. The multi-head design of MACE-MH-1 partly mitigates this, yet the correct head should still be selected explicitly for a single-functional target.

Long-range and polarisable systems are out of scope for most weights. Most MACE weights express short-range truncated local energies, which is inadequate for polar liquids, ionic systems and interfacial double layers where explicit long-range electrostatics matter. MACE-POLAR-1 introduces a non-self-consistent polarisable field for exactly this reason, but covers 83 elements on one dataset, and the test is direct: if the electrostatic contribution remains significant beyond the cutoff radius, standard weights do not apply.

Vibrational spectra and high-pressure phases carry known defects. The foundations README documents the generation-by-generation corrections, with the 0b release improving pair repulsion and isolated atom energies, 0b2 improving stability at high pressure, and 0b3 fixing some phonon issues relative to 0b2. That sequence itself indicates that phonons and high-pressure properties are the weak spot of machine-learned potentials and that they were still being corrected. Users concerned with phonons or high-pressure behaviour should prefer the newest generation and run density-functional-theory spot checks on the critical results.

Ground-state energy conventions can hide thermodynamic errors. Because of how isolated-atom energy offsets are handled, formation energies computed from a potential depend on the reference-state convention in the training data. Mixing energies produced by different batches or different models when building phase diagrams or running thermodynamic integration introduces a systematic error that is hard to notice. The mitigation is to generate all energies from one model under one offset convention, and to calibrate against density-functional-theory samples before any phase-stability comparison.

The v1.0 rewrite introduces breakage. The README states that MACE is being rewritten toward v1.0, that the rewrite happens in the open on a dedicated branch with a public progress board, and that v0.3.x is the supported production line. v1.0 will not preserve code compatibility, and trained checkpoints will migrate through an explicit converter rather than loading directly. The practical consequence is that long-lived projects should pin a v0.3.x version, track the deprecation list in the release notes, and avoid switching to the new main line before v1.0 settles.

Licence risk is not a technical failure mode, but it fails the same way. The first-generation MP weights are MIT, whereas OMAT, MATPES, MH, OMOL, POLAR-1 and MACE-OFF23 are under the Academic Software License and restricted to academic use. Using Academic-Software-License weights inside a commercial product or in externally delivered work constitutes a breach of the licence, and this must be settled at tool-selection time rather than discovered before release.

9. Version and Source Notes

Versions need to be recorded in layers, because the word means different things at different levels.

For code, the main repository has 18 releases, beginning with the v0.2.0 alpha of 9 February 2023 and continuing through v0.3.0 in November 2023, v0.3.2 in December 2023, v0.3.3 and v0.3.4 in January 2024, v0.3.5 in June 2024, v0.3.6 in July 2024, v0.3.7 in October 2024, v0.3.8 and v0.3.9 in November 2024, v0.3.10 in February 2025, v0.3.11 and v0.3.12 in March 2025, v0.3.13 in April 2025, v0.3.14 in August 2025, v0.3.15 in February 2026, v0.3.16 in May 2026 and v0.3.17 in October 2026. There are 21 Git tags in total, including non-release tags such as v0.3.7b. The PyPI distribution mace-torch tracks the GitHub releases, with 15 releases and the current 0.3.17 uploaded on 27 October 2026.

For weights, the ten releases of the foundations repository correspond to the model generations described earlier, while an earlier dedicated repository carries the first generation.

For archiving, the Zenodo DOI 10.5281/zenodo.14103332 resolves through DataCite to the record titled ACEsuit/mace: v0.3.17, of resource type Software, published in 2026, with a CC-BY-4.0 rights statement and 30 creators. Supplementary materials for the foundation models live in the Zenodo community for MACE.

On branches, the main repository default branch is develop rather than main, and the v1.0 rewrite sits on its own branch, while the foundations, off and JAX repositories default to main. A plain clone therefore lands on the development line, so a stable checkout must name a v0.3.x tag explicitly.

Three caveats on citation are worth recording. The GitHub API reports the licence of the main repository as unasserted, because the licence file is named LICENSE.md and the format is not auto-detected; the actual licence is standard MIT, and the README badge reading MIT 2.0 is inaccurate and should not be relied on in place of the LICENSE.md text. The preprint and journal versions of each paper must be cited separately, since the 2023 foundation-model paper has an arXiv version and a 2025 journal version, and MACE-OFF23 has an arXiv preprint and a 2025 journal article in the Journal of the American Chemical Society.

10. References

  1. MACE code repository and release page. https://github.com/ACEsuit/mace , DOI: https://doi.org/10.5281/zenodo.14103332
  2. Batatia I., Kovacs D.P., et al. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. Advances in Neural Information Processing Systems 35, 11423-11436, 2022. DOI: https://doi.org/10.52202/068431-0830
  3. Batatia I., Benner P., Chiang Y., et al. A foundation model for atomistic materials chemistry. arXiv:2401.00096, 2023. DOI: https://doi.org/10.48550/arXiv.2401.00096
  4. Batatia I., Benner P., Chiang Y., et al. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics, Vol. 163, 2025. DOI: https://doi.org/10.1063/5.0297006
  5. MACE pretrained foundation model weights repository. https://github.com/ACEsuit/mace-foundations
  6. MACE-OFF23 transferable organic force field model repository. https://github.com/ACEsuit/mace-off
  7. Kovacs D.P., Moore J.H., Browning N.J., et al. MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules. Journal of the American Chemical Society, 2025. DOI: https://doi.org/10.1021/jacs.4c07099
  8. Kovacs D.P., Moore J.H., Browning N.J., et al. MACE-OFF23: Transferable Machine Learning Force Fields for Organic Molecules. arXiv:2312.15211, 2023. DOI: https://doi.org/10.48550/arXiv.2312.15211
  9. MACE-OFF23 training and evaluation dataset. Cambridge Data Repository, 2024. DOI: https://doi.org/10.17863/CAM.107498
  10. Riebesell J., Goodall R.E.A., Benner P., et al. Matbench Discovery - A framework to evaluate machine learning crystal stability predictions. arXiv:2308.14920, 2023. DOI: https://doi.org/10.48550/arXiv.2308.14920
  11. Yuan E.C.-Y., et al. Foundation models for atomistic simulation of chemistry and materials. Nature Reviews Chemistry, 10, 212-230, 2026. DOI: https://doi.org/10.1038/s41570-025-00793-5
  12. Zhang S., et al. Node-equivariant message passing for efficient and accurate machine learning interatomic potentials. Chemical Science, 17, 3793-3803, 2026. DOI: https://doi.org/10.1039/d5sc07248d
  13. Deng B., Zhong P., Jun K., et al. CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling. arXiv:2302.14231, 2023. DOI: https://doi.org/10.48550/arXiv.2302.14231
  14. Gonnheimer M., et al. MACE-MDP: A Foundation Model for Molecular Dipole Moments and Polarizabilities. ChemRxiv, 2026. DOI: https://doi.org/10.26434/chemrxiv.15000716/v1
  15. MACE official documentation. https://mace-docs.readthedocs.io/
  16. MACE PyPI distribution package mace-torch. https://pypi.org/project/mace-torch/
  17. MACE tutorial repository. https://github.com/ACEsuit/mace-tutorials
  18. ACE.jl equivariant computation foundation package. https://github.com/ACEsuit/ACE.jl
Tags: