MACE Neural Network Potential

• MACE Neural Network Potentials are rotation-equivariant, message-passing force fields that embed atomic cluster expansion into a deep neural architecture for accurate atomistic modeling.
• They capture high-order many-body correlations through tensorized local descriptors and strict O(3) symmetry, enabling precise representation of molecular and material interactions.
• Benchmark studies demonstrate that MACE outperforms conventional ML force fields with superior accuracy, efficiency, and transferability across diverse systems.

MACE (Multilayer Atomic Cluster Expansion) neural network potentials are a class of rotation-equivariant, message-passing machine learning interatomic potentials that generalize the atomic cluster expansion (ACE) formalism by embedding it in a deep neural architecture. MACE models achieve state-of-the-art accuracy and transferability in atomistic modeling across organic molecules, condensed-phase systems, and materials, with efficient scaling and superior low-data learning compared to conventional force fields and lower-body-order message-passing neural networks (Bernstein, 2024, Batatia et al., 2022, Kovács et al., 2023, Kovacs et al., 2023).

1. Theoretical Foundations and Architecture

MACE unifies three developments in machine-learned interatomic potentials: nonlinear neural networks, high body-order expansions, and strict $O(3)$ equivariance. At its core, MACE uses local atomic environments encoded by generalized ACE descriptors up to a chosen body order $\nu$ (typically $\nu=4$), which fully parameterize many-body correlations within a cutoff radius. Each atom $i$ is a node in a molecular or crystalline graph, connected to neighbors $j$ via edges encoding radial ($R_n(r_{ij})$) and angular ($Y_\ell^m(\hat r_{ij})$) information, with chemical identity handled by Kronecker deltas or learned embeddings (Bernstein, 2024, Kovács et al., 2023).

The MACE network consists of multiple message-passing layers. At each layer $t$, atom-wise features $h_i^{(t)}$ are iteratively updated via equivariant aggregation of symmetrized tensor products of local descriptors from neighboring atoms, using small multilayer perceptrons (MLPs) for message construction and updates. Messages

$$m_{i}^{(t)} = \sum_{j\in N(i)} \mathrm{MLP}^{(t)}(B_i, B_j)$$

are combined and added to $h_i^{(t)}$ in a residual fashion:

$h_{i}^{(t+1)} = h_{i}^{(t)} + U^{(t)} m_{i}^{(t)}$

with $U^{(t)}$ a learnable linear map. The network is typically very shallow: $T=2$ layers suffice, because each layer handles up to $\nu+1$-body terms via tensorized products, efficiently capturing high-order interactions without deep stacking (Batatia et al., 2022, Bernstein, 2024).

After message passing, a final readout block (MLP) maps the last-layer features $h_i^{(T)}$ to an atomic energy contribution $\epsilon_i$. The total potential energy is the sum over sites:

$E = \sum_i \epsilon_i(h_i^{(T)})$

Forces and stresses are obtained by analytic differentiation with respect to atomic positions and strain, ensuring energy conservation and compatibility with molecular dynamics (Bernstein, 2024, Batatia et al., 2022, Kovács et al., 2023).

2. Descriptor Construction and Equivariance

MACE employs descriptors rooted in the ACE formalism:

• Radial basis: $R_n(r)$, e.g., orthonormal polynomials or Gaussians times a smooth cutoff $f_{\rm cut}(r)$; cutoff radii $R_c$ are typically 5–12 Å, enforcing strict locality.
• Angular basis: Real spherical harmonics $Y_\ell^m(\theta, \phi)$ up to $\ell_{\max}$ (usually $\leq 2$ or 3), encoding angular correlations.
• Chemical embedding: One-hot or linear-embedded delta functions $\delta_{z_j,z_u}$ reduce combinatorial growth for multicomponent systems.
• Body-order truncation: Tensor products of the atomic basis yield site features up to $\nu$-body, recoupled via Clebsch–Gordan coefficients into overall tensor rank $\lambda$, ensuring $SO(3)$ rotation symmetry.

These descriptors enable contracting local geometric and elemental information to arbitrary body order, maximally exploiting local correlation structure while keeping input dimensionality tractable even for systems with up to 89 elements (Bernstein, 2024, Shiota et al., 2024, Kovács et al., 2023).

Strict $O(3)$ equivariance is maintained at every layer: all features and messages transform as irreducible representations under spatial rotations, enabling exact encoding of angular dependencies in the potential and ensuring physical consistency under rigid-body motions.

3. Training Methodologies and Hyperparameters

MACE models are trained via supervised regression to electronic structure reference data, fitting total energies, atomic forces, and optionally stress components. The loss function is a weighted sum, often with force terms dominant to enforce good gradients:

$$L = w_E \sum_k (E^{\mathrm{pred}}_k - E^{\mathrm{ref}}_k)^2 + w_F \sum_{k,i} \|\mathbf{F}^{\mathrm{pred}}_{k,i} - \mathbf{F}^{\mathrm{ref}}_{k,i}\|^2 + w_\sigma \sum_k \|\sigma^{\mathrm{pred}}_k - \sigma^{\mathrm{ref}}_k\|^2$$

Huber losses and "conditional" force thresholds are sometimes used to prioritize robustness (Alghamdi et al., 6 Oct 2025). Weight decay and learning-rate scheduling are standard; "multi-head stabilization," i.e., attaching auxiliary readouts to intermediate layers, is used during fine-tuning to improve stability. Training employs Adam or AdamW optimizers at initial rates $10^{-3}$ to $10^{-4}$, typically for $10^5$ steps, with batch sizes from 1 to 32 depending on memory (Bernstein, 2024, Kovács et al., 2023, Alghamdi et al., 6 Oct 2025, Park et al., 24 Mar 2025).

Hyperparameters such as message-passing depth ($T=2$), body order ($\nu=4$), channel widths (32–256), and radial basis size (6–32) can be tuned for the application domain, balancing accuracy, cost, and memory.

4. Benchmark Performance and Applications

MACE outperforms or matches leading MLIP architectures (GAP, ACE, DeePMD, NequIP, M3GNet) on established molecular and material benchmarks. Representative results include:

Model	Energy MAE [meV/atom]	Force MAE [meV/Å]	Stress MAE [meV/Å$^3$]	Speed [ms/atom/step]
GAP (SOAP, $R_c=6.5$ Å)	8.2	78	3.0	2.1 (CPU)
ACE (body-order 5)	8.5	82	2.4	0.18 (CPU)
MACE-small	5.8	76	1.7	0.042 (GPU) (LAMMPS)
MACE-medium	5.7	80	1.7	0.12 (GPU)

In molecular benchmarks, MACE achieves sub-5–10 meV/atom accuracy with two message-passing layers and a $\sim$10 Å receptive field. In organic-chemical and biomolecular benchmarks (MACE-OFF23), test MAEs reach as low as 0.9–1.5 meV/atom and 14–36 meV/Å for forces, with transferability across small molecules, crystals, liquids, and folded proteins (Kovács et al., 2023, Kovacs et al., 2023).

Fine-tuning foundation models (e.g., MACE-MP0 for general chemistry or MACE-OFF23 for organics) with a few hundred data points enables rapid adaptation to new systems with performance on par with models trained on orders of magnitude more data. For battery-relevant LiF diffusion, activation barriers and diffusivities predicted by MACE-MP0 or fine-tuned variants agree within 10% of DeePMD models trained on $>40,000$ active-learning configurations, using only $<300$ DFT frames (Alghamdi et al., 6 Oct 2025).

MACE has also demonstrated transferability in “out-of-the-box” NMR shift prediction (as descriptor for kernel regression) and liquid properties such as density and diffusion for ionic liquids, systematically outperforming or matching alternatives and empirical force fields (Shiota et al., 2024, Park et al., 24 Mar 2025).

5. Specialized Variants and Extensions

Significant domain extensions include:

• MACE-OFF23: Specialized for neutral organic chemistry, trained on SPICE and QMugs datasets, captures condensed-phase and biopolymer behavior without explicit long-range corrections.
• X-MACE: Incorporates DeepSets-based invariant encoding to model non-smooth excited-state potential energy surfaces, such as near conical intersections, with large accuracy gains and robust transfer learning from MACE ground-state models (Barrett et al., 18 Feb 2025).
• Foundation models: Large-scale pretraining on diverse datasets (Materials Project, Alexandria) yields universal "MP0" or "OFF23" potentials that can be efficiently fine-tuned for rapid deployment to specific chemical systems, including ab initio thermochemistry, reactive MD, or spectroscopy at CCSD(T) or DFT levels (Alghamdi et al., 6 Oct 2025, Shiota et al., 2024).

6. Transferability, Locality, and Computational Efficiency

The strictly local, body-ordered construction of MACE ensures rapid evaluation ($0.04$–$0.12$ ms/atom/step on GPU for $R_c\sim 10$ Å) and transferability to large, complex systems spanning different elements and bonding motifs. Locality up to $2 \times R_c$ is typically sufficient for most chemical and van der Waals interactions. Large-biopolymer simulations (e.g., solvated proteins $>10^4$ atoms) are supported, with performance limited primarily by GPU memory and lack of multi-GPU domain decomposition in current implementations (Kovács et al., 2023, Bernstein, 2024).

Long-range electrostatics and explicit dispersion tails are not included by construction, but learned many-body expansions can implicitly recover such effects within the cutoff. For highly charged or multipolar systems, future development may incorporate explicit long-range terms (Kovacs et al., 2023).

7. Limitations, Best Practices, and Outlook

MACE offers robust, accurate, and transferable MLIPs with high data efficiency but entails higher training complexity and memory cost than linear ACE or GAP. Best practices include utilization of pretrained foundation models, multi-head stabilization during fine-tuning, weight decay, learning-rate scheduling, and careful choice of cutoff and body order. Current models are restricted to single-GPU or limited parallelism; scaling to $\mathcal{O}(10^5)$ atoms and multi-GPU/CPU implementations is an active area of research (Bernstein, 2024, Kovács et al., 2023).

MACE’s combination of linearly complete many-body descriptors and equivariant deep learning positions it as a leading paradigm for next-generation machine-learned force fields in chemistry and materials science, bridging molecules, solids, and complex interfaces with minimal tuning or data requirements.

---

Key References

• N. Bernstein, "From GAP to ACE to MACE" (Bernstein, 2024)
• I. Batatia et al., "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields" (Batatia et al., 2022)
• I. Batatia et al., "MACE-OFF: Transferable Short Range Machine Learning Force Fields for Organic Molecules" (Kovács et al., 2023)
• Shiota et al., "Universal neural network potentials as descriptors..." (Shiota et al., 2024)
• R. Schmid et al., "Transferable Machine Learning Potential X-MACE for Excited States using Integrated DeepSets" (Barrett et al., 18 Feb 2025)
• Y. Park et al., "Ionic Liquid Molecular Dynamics Simulation with Machine Learning Force Fields: DPMD and MACE" (Park et al., 24 Mar 2025)
• C. F. Anderson et al., "Evaluation of the MACE Force Field Architecture..." (Kovacs et al., 2023)
• R. Car, "Comparing fine-tuning strategies of MACE machine learning force field for modeling Li-ion diffusion in LiF for batteries" (Alghamdi et al., 6 Oct 2025)