MACE

MACE systematically incorporates many-body interactions beyond the typical three-body terms used in other potentials. By leveraging the Atomic Cluster Expansion (ACE) formalism, it constructs a complete basis of body-ordered polynomials. This allows the model to explicitly capture four-body, five-body, and higher-order atomic correlations that are critical for describing complex chemical environments, such as hydrogen bonding networks and transition states, without relying on implicit, deep network memorization.

The architecture operates on irreducible representations (irreps) of the O(3) rotation group, ensuring strict E(3) equivariance. Messages between atoms are constructed using tensor products of the spherical harmonic projections of interatomic distance vectors with neighboring node features. This guarantees that the predicted potential energy is invariant to translation and rotation, while internal equivariant features transform correctly, preserving directional information like forces and multipoles without data augmentation.

MACE achieves high computational efficiency by decoupling the body-order expansion from the depth of the message-passing network. A single MACE layer can construct high body-order features internally, allowing the model to be remarkably shallow—often requiring only two message-passing layers to achieve state-of-the-art accuracy. This contrasts with deep GNNs that require many layers to capture long-range many-body effects, resulting in significantly faster inference and reduced memory consumption for large-scale molecular dynamics simulations.

Forces are computed analytically via exact differentiation of the energy output with respect to atomic positions, ensuring energy conservation. The final energy is obtained by summing invariant, scalar outputs from a readout function applied to the final node features. Because the internal features are equivariant, the model naturally outputs vector-valued forces that rotate correctly with the molecule, making it directly suitable for running stable, long-time-scale molecular dynamics simulations without iterative coordinate alignment.

By embedding the physical prior of many-body equivariance directly into the architecture, MACE exhibits exceptional data efficiency. It learns accurate potential energy surfaces from significantly smaller training datasets compared to invariant models or simpler equivariant architectures. This is crucial for high-cost quantum mechanical reference data, such as CCSD(T) calculations, where generating large training sets is computationally prohibitive. The model's inductive bias allows it to generalize accurately from sparse reference calculations.

Recent developments have scaled MACE into a universal interatomic potential (MACE-MP-0) trained on the Materials Project database covering 89 elements. This single model accurately predicts energies, forces, and phonon properties across a vast swath of the periodic table without retraining. It serves as a general-purpose, transferable force field for high-throughput screening of inorganic materials, crystal structure prediction, and defect analysis, rivaling the accuracy of specialized single-system potentials.

An E(3)-equivariant neural network interatomic potential that directly preceded MACE, using tensor products of irreducible representations to achieve data-efficient force predictions. NequIP demonstrated that equivariance to 3D rotations and translations dramatically reduces the training data required for accurate potential energy surfaces. It operates on spherical harmonic embeddings of atomic environments but uses a simpler message-passing scheme compared to MACE's higher-order many-body expansion.

A systematic and complete basis set expansion of atomic environments that forms the theoretical backbone of MACE. ACE constructs body-ordered invariant features by projecting atomic densities onto a hierarchical polynomial basis, yielding highly efficient linear models. MACE extends ACE by embedding this expansion within a message-passing neural network framework, using higher-order tensor products to capture many-body interactions beyond the three-body terms typical of simpler potentials.

A fundamental geometric constraint requiring that a function's output transforms consistently with any input rotation and translation in 3D Euclidean space. For molecular potentials, this means rotating a molecule produces identically rotated forces. MACE enforces strict SE(3) equivariance through Clebsch-Gordan tensor products of irreducible representations, ensuring physical consistency without data augmentation. This property is critical for conservation of linear and angular momentum in molecular dynamics simulations.

The general framework underlying MACE's architecture, where node representations are iteratively updated by aggregating information from neighboring nodes via message and update functions. MACE extends standard MPNNs by constructing messages using equivariant many-body features rather than simple scalar distances. Each message incorporates higher-order tensor products of neighboring atomic environments, allowing the model to capture complex angular dependencies essential for accurate potential energy prediction.

A transformer architecture that integrates SE(3)/E(3) equivariance using tensor products and equivariant attention mechanisms. Equiformer applies attention weights to equivariant features rather than scalar features alone, achieving state-of-the-art performance on 3D molecular property prediction benchmarks. While MACE focuses on interatomic potentials with many-body message construction, Equiformer demonstrates how attention can be made equivariant for broader molecular tasks.

A pioneering continuous-filter convolutional neural network that models quantum interactions by using interatomic distances to generate filter kernels for message passing. SchNet introduced the concept of continuous spatial convolutions on molecular graphs, where filter functions depend smoothly on pairwise distances. While SchNet is invariant rather than equivariant and lacks the many-body tensor product structure of MACE, it established the foundational paradigm of learning from atomic coordinates directly.