MACE-H Graph Neural Network for Hamiltonian Prediction

• The paper introduces MACE-H, which accurately predicts Kohn–Sham Hamiltonians by combining high body-order message passing with a node-order expansion tailored for orbital-resolved interactions.
• It employs equivariant features and Clebsch–Gordan projections to generate all necessary O(3) irreducible representations for s, p, d, and f orbital blocks.
• The model achieves sub-meV prediction errors on matrix elements across diverse systems, highlighting its efficiency for high-throughput electronic-structure screening.

to արտաքին? No, final only. Searching arXiv for the cited papers to ground the article in the current literature. MACE-H is an $O(3)$-equivariant many-body graph neural network for predicting Kohn–Sham density functional theory Hamiltonians from atomistic structure. In the formulation introduced in “Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing,” MACE-H combines high body-order message passing with a node-order expansion so that the model can generate all relevant $O(3)$ irreducible representations needed for block-structured Hamiltonians up to $f$ orbitals, including on-site and inter-site interactions and, where relevant, spin-resolved channels under spin–orbit coupling (SOC) (Qian et al., 20 Aug 2025). The name also sits within a broader lineage: MACE itself originates from higher-order equivariant message passing in interatomic force-field learning (Batatia et al., 2022), while a later review of GAP, ACE, and MACE noted that “MACE-H” was not yet a separately defined architecture in that context and instead connected the label to hierarchical or multihead stabilizing modifications used during fine-tuning (Bernstein, 2024). As a result, the term has both a precise meaning in Hamiltonian learning and an earlier, looser usage in the wider MACE literature.

1. Nomenclature and historical placement

The immediate antecedent of MACE-H is MACE, introduced as a “Higher Order Equivariant Message Passing Neural Network” for fast and accurate force fields (Batatia et al., 2022). That work argued that most equivariant message passing neural networks rely on two-body message passing, so higher-body correlations must be built indirectly through depth and nonlinearities. MACE addressed this by constructing higher body-order messages, specifically showing that “using four-body messages reduces the required number of message passing iterations to just two” (Batatia et al., 2022). In that original force-field setting, however, the term “MACE-H” was not defined as a separate model.

A later review, “From GAP to ACE to MACE,” placed MACE within the evolution from SOAP–GAP to ACE and then to equivariant message-passing models, but it explicitly stated that the review “does not introduce or define ‘MACE-H’ as a separate architecture” (Bernstein, 2024). In that review’s usage, the closest related idea was “fine-tuning (with multihead stabilization),” together with broader hierarchical or hybrid design choices.

The specific architecture now called MACE-H appears in the Hamiltonian-learning literature. In “Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing,” MACE-H is defined as a graph neural network that extends the Many-body Atomic Cluster Expansion paradigm from energies and forces to electronic Hamiltonians (Qian et al., 20 Aug 2025). This usage is distinct from ACE-HGNN, the “Adaptive Curvature Exploration Hyperbolic Graph Neural Network,” which is a multi-agent reinforcement-learning framework for hyperbolic representation learning and does not use the string “MACE-H” in the paper itself (Fu et al., 2021).

This naming history matters because it prevents a common misconception: MACE-H is not a single universally fixed acronym across all graph-learning subfields. In electronic-structure prediction, it denotes a concrete Hamiltonian model; in earlier MACE discussions, it could only be inferred as shorthand for higher-order, hierarchical, or multihead variants.

2. Architectural definition in Hamiltonian learning

In its explicit 2025 form, MACE-H is designed to predict Hamiltonian matrix elements $H_{ij}$ between localized orbitals, including on-site blocks $(i=j)$ and inter-site blocks $(i\neq j)$, with orbital shells spanning $s$ $(\ell=0)$, $p$ $(\ell=1)$, $O(3)$0 $O(3)$1, and $O(3)$2 $O(3)$3 channels (Qian et al., 20 Aug 2025). The core architectural statement is that it “combines high body-order message passing with a node-order expansion to efficiently obtain all relevant $O(3)$4 irreducible representations” (Qian et al., 20 Aug 2025).

The graph construction is standard for atomistic GNNs. Nodes are atoms; edges connect neighbors within a radial cutoff; inputs include atomic species $O(3)$5, positions $O(3)$6, lattice vectors, and periodic boundary conditions. Node features carry species embeddings and equivariant hidden states arranged by irreducible representations up to a chosen $O(3)$7, while edge features decompose relative geometry into radial basis functions and spherical harmonics $O(3)$8 (Qian et al., 20 Aug 2025).

For each node $O(3)$9 and angular momentum $f$0, equivariant features are tensors $f$1 with components $f$2. On an edge $f$3, the model uses a radial–angular decomposition

$f$4

with learnable coefficients $f$5 that may be species-dependent (Qian et al., 20 Aug 2025). Equivariant couplings are implemented through Clebsch–Gordan decomposition, so tensor products of irreps are projected into target irreps with the correct $f$6 transformation law.

A prototypical update is written as

$f$7

followed by an equivariant nonlinearity (Qian et al., 20 Aug 2025). Because the outputs are Hamiltonian blocks rather than scalar energies, equivariance is not a decorative property but a structural requirement: $f$8, $f$9, and $H_{ij}$0 orbital blocks are intrinsically direction-sensitive.

The output stage maps equivariant node and edge features to shell-resolved blocks $H_{ij}$1. Hermiticity is imposed at assembly time through $H_{ij}$2, and on-site blocks are symmetrized (Qian et al., 20 Aug 2025). For periodic systems, real-space blocks $H_{ij}$3 are Fourier-summed into

$H_{ij}$4

which enables band-structure and density-of-states calculations from predicted matrix elements (Qian et al., 20 Aug 2025).

3. High body-order message passing and node-order expansion

The many-body component of MACE-H is inherited conceptually from MACE. The original MACE paper framed its central advance as the use of higher body-order messages so that body order is no longer tied to network depth; by constructing four-body messages, MACE reaches high expressivity with only two message-passing iterations (Batatia et al., 2022). The later GAP–ACE–MACE review likewise described MACE as feeding ACE’s equivariant and invariant tensor bases into a message-passing graph neural network that is explicitly $H_{ij}$5-equivariant (Bernstein, 2024).

MACE-H adapts this logic to Hamiltonian prediction. High body-order message passing is defined through symmetric polynomials over neighbor contributions, projected to target angular momenta by Clebsch–Gordan decomposition:

$H_{ij}$6

with $H_{ij}$7 (Qian et al., 20 Aug 2025). These terms create effective $H_{ij}$8-body features without naïve explicit enumeration of all triplets or quadruplets.

The node-order expansion is the second defining ingredient. To ensure that hidden features can parameterize Hamiltonian blocks up to $H_{ij}$9 orbitals, MACE-H forms higher-order irreps by tensor products of node features followed by Clebsch–Gordan projection:

$(i=j)$0

This “bootstraps richer angular content from learned lower-$(i=j)$1 channels” and is the mechanism by which the model can efficiently obtain all relevant irreps for $(i=j)$2 Hamiltonian blocks (Qian et al., 20 Aug 2025).

A plausible implication is that the phrase “MACE-H” in the Hamiltonian context can be read as an extension of the original “higher order” idea from force-field MACE into a representation space tailored to orbital-block prediction. That interpretation is consistent with the historical MACE emphasis on body order (Batatia et al., 2022), but the Hamiltonian paper’s distinctive addition is the node-order expansion (Qian et al., 20 Aug 2025).

4. Physical constraints, training objective, and stabilization

MACE-H is trained primarily on matrix-element prediction. A canonical multi-task objective given in the Hamiltonian paper is

$(i=j)$3

where $(i=j)$4 denotes eigenvalues and $(i=j)$5 collects regularization and symmetry penalties (Qian et al., 20 Aug 2025). In the reported experiments, the primary metrics are mean absolute errors of matrix elements, while eigenvalues are validated post hoc for bulk Au (Qian et al., 20 Aug 2025).

The paper emphasizes a practical difficulty specific to Hamiltonian learning under SOC: different element–orbital pair sub-blocks can vary in magnitude by orders, which causes gradient instability under naïve scaling. To address this, MACE-H introduces a “shift-and-scale” operation for the last-layer irreps,

$(i=j)$6

where $(i=j)$7 and $(i=j)$8 are precomputed means and standard deviations for each sub-block type, and the detach operation prevents the scaling factors from amplifying backpropagated gradients (Qian et al., 20 Aug 2025). The reported effect is a reduction in mean absolute error and improved convergence, especially for small-magnitude cross-spin blocks.

Several physical constraints are built into the workflow rather than left entirely to optimization. Hermiticity is enforced by symmetrization; time-reversal symmetry and permutation invariance are respected by construction of features and aggregation; periodic assembly is handled through real-space-to-$(i=j)$9-space Fourier summation (Qian et al., 20 Aug 2025). For bulk Au, the paper further reports that core-state projection and sparsification beyond $(i\neq j)$0 Å produce negligible effects on eigenvalues, with eigenvalue error $(i\neq j)$1 eV and entropy error $(i\neq j)$2 eV K$(i\neq j)$3 Å$(i\neq j)$4 relative to model error (Qian et al., 20 Aug 2025).

The broader MACE review provides context for why these stabilization devices are noteworthy. In that review, multihead stabilization was highlighted as a practically important modification during fine-tuning of foundation models, intended to improve optimization robustness (Bernstein, 2024). Although that is not the same mechanism as the Hamiltonian paper’s shift-and-scale operation, both address the same general issue: training equivariant many-body models can be limited by optimization stability rather than representational capacity alone.

5. Datasets, reported performance, and empirical behavior

The Hamiltonian paper evaluates MACE-H on several open materials benchmark datasets and a new dataset for bulk gold (Qian et al., 20 Aug 2025). The 2D datasets include monolayer graphene, monolayer MoS$(i\neq j)$5, bilayer graphene, bilayer Bismuthene, bilayer Bi$(i\neq j)$6Se$(i\neq j)$7, and bilayer Bi$(i\neq j)$8Te$(i\neq j)$9, with SOC included for Bismuthene, Bi$s$0Se$s$1, and one Bi$s$2Te$s$3 setting (Qian et al., 20 Aug 2025). Bulk Au is computed with FHI-aims and treated without SOC (Qian et al., 20 Aug 2025).

The paper reports the following representative matrix-element MAEs in meV:

System	Baseline(s)	MACE-H
Graphene (noSOC) monolayer	DeepH 2.1; DeepH-E3 0.27	0.21
MoS$s$4 (noSOC) monolayer	DeepH 0.95; DeepH-E3 0.45	0.39
Bilayer graphene (noSOC), shifted/twisted	DeepH-E3 0.40/0.28	0.31/0.24
Bilayer Bi$s$5Se$s$6 (SOC), shifted/twisted	DeepH-E3 0.40/0.35	0.34/0.35
Bulk Au (noSOC)	DeepH-E3 0.41 (max 0.80)	0.32 (max 0.65)
Bulk Au + shift-and-scale	—	0.27 (max 0.62)

Across monolayers and shifted bilayers, MACE-H is reported to outperform DeepH-E3 and non-equivariant DeepH, with overall MAEs “down to $s$7–$s$8 meV” (Qian et al., 20 Aug 2025). The same paper also reports that MACE-H “consistently improves in-distribution shifted bilayers,” whereas “out-of-distribution twisted bilayers are more challenging for high $s$9 and high $(\ell=0)$0” (Qian et al., 20 Aug 2025). This is not a contradiction; it reflects a stated locality preference.

For bulk Au, the reference model uses $(\ell=0)$1 layers, $(\ell=0)$2, $(\ell=0)$3, $(\ell=0)$4, $(\ell=0)$5, learning rate $(\ell=0)$6, and batch size $(\ell=0)$7 (Qian et al., 20 Aug 2025). For monolayer graphene, an example configuration is learning rate $(\ell=0)$8, batch size $(\ell=0)$9, $p$0 layers, $p$1, and $p$2 (Qian et al., 20 Aug 2025). These concrete settings indicate that the model is not tied to a single hyperparameter regime.

The original MACE paper provides a complementary empirical baseline for the higher-order design principle. On rMD17, 3BPA, and AcAc, MACE reached or exceeded state-of-the-art accuracy while remaining substantially faster than NequIP or BOTNet at comparable accuracy in several settings (Batatia et al., 2022). This does not establish Hamiltonian performance directly, but it clarifies why higher body-order message passing was a natural foundation for extending MACE toward more demanding equivariant targets.

6. Computational properties, limitations, and relation to adjacent models

The Hamiltonian paper states that message passing per layer scales roughly as

$p$3

where $p$4 is the number of atoms, $p$5 the number of neighbors per atom, and $p$6 the number of channels for irrep $p$7 (Qian et al., 20 Aug 2025). Many-body terms add polynomial factors up to order $p$8, but these are implemented via tensor factorizations and Clebsch–Gordan projections so as to avoid the combinatorial blow-up of naïve $p$9-body enumeration (Qian et al., 20 Aug 2025). In the reported settings, inference time on a single A100 GPU is comparable to DeepH-E3 because the dominant cost lies in the shared edge-update block, while the many-body expansion and node-order expansion add negligible overhead (Qian et al., 20 Aug 2025).

The same work identifies a central trade-off between locality and transferability. Higher $(\ell=1)$0 and higher $(\ell=1)$1 improve local shifted-bilayer predictions but can worsen twisted-bilayer out-of-distribution performance; for twisted systems, the recommendation is to reduce $(\ell=1)$2 and $(\ell=1)$3 and enlarge the training set (Qian et al., 20 Aug 2025). Perturbation-response analysis further shows that MACE-H exhibits “faster attenuation with distance,” which the authors interpret as consistent with its locality preference (Qian et al., 20 Aug 2025).

This locality issue aligns with a broader limitation already recognized in MACE and ACE more generally. The 2024 review noted that short-range ACE/MACE frameworks do not by themselves solve long-range electrostatics or charge transfer, and that improved scaling and parallelism remain open engineering directions (Bernstein, 2024). The Hamiltonian paper formulates the corresponding limitation in electronic-structure language: long-range interactions beyond the cutoff may be underrepresented, and nonlocal terms may be needed for moiré or twist systems (Qian et al., 20 Aug 2025).

Comparison with neighboring architectures is therefore best made along two axes. Relative to DeepH and DeepH-E3, MACE-H adds explicit many-body and node-order machinery while preserving $(\ell=1)$4 equivariance (Qian et al., 20 Aug 2025). Relative to the original force-field MACE, it changes the target from scalar site energies and derived forces to block-structured Hamiltonians with shell-resolved angular content (Batatia et al., 2022). A common misconception is to treat MACE-H as merely “MACE plus a new output head.” The Hamiltonian paper suggests a stronger claim: the node-order expansion is specifically introduced to generate all irreps needed for $(\ell=1)$5-orbital matrix interaction blocks, which is a target-space requirement absent from standard interatomic-potential formulations (Qian et al., 20 Aug 2025).

A final source of confusion comes from unrelated terminology in graph learning. ACE-HGNN is a hyperbolic graph neural network with adaptive curvature exploration via multi-agent reinforcement learning (Fu et al., 2021). Despite superficial overlap in initials, it belongs to a different methodological lineage and addresses a different problem.

7. Significance and scope of current usage

Within electronic-structure machine learning, MACE-H occupies the intersection of equivariant GNNs, ACE-style many-body expansions, and Hamiltonian surrogate modeling. Its stated objective is to enable rapid prediction of band structures and densities of states without repeated self-consistent DFT, by learning localized Hamiltonian blocks and then assembling periodic Hamiltonians in $(\ell=1)$6-space (Qian et al., 20 Aug 2025). The reported outcome is “sub-meV prediction errors on matrix elements and eigenvalues across all systems,” together with computational efficiency suitable for high-throughput screening (Qian et al., 20 Aug 2025).

In the longer genealogy of atomistic machine learning, MACE-H can be viewed as a specialization of the MACE design philosophy rather than a replacement of it. GAP supplied symmetry-correct descriptors and probabilistic regression; ACE supplied a systematic symmetry-adapted basis with controllable body order; MACE fused ACE-like equivariant structure with message passing; and MACE-H extends that route to orbital-resolved Hamiltonian prediction (Bernstein, 2024). This suggests a broader pattern: once equivariant many-body features are available, the main remaining question is how to tailor the representation to the physical tensorial object being predicted.

The term “MACE-H Graph Neural Network” therefore has a precise modern referent: an $(\ell=1)$7-equivariant many-body GNN for Kohn–Sham Hamiltonians, built from high body-order message passing and node-order expansion (Qian et al., 20 Aug 2025). At the same time, the historical record shows that earlier MACE literature used adjacent ideas—higher-order message passing, hierarchical aggregation, and multihead stabilization—without defining a separate MACE-H architecture (Batatia et al., 2022, Bernstein, 2024). Recognizing that distinction is essential for accurate reading of the recent literature.