Equivariant Cellular Sheaves for Molecular Electronic Structure: Bridging Sheaf Cohomology and E(3)-Equivariant Hamiltonian Learning

• A shifted localized-orbital Hamiltonian becomes a cellular-sheaf Laplacian on a molecular cell complex, connecting geometry, chemical topology, and electronic spectra.The resulting model is called the Equivariant Cellular Sheaf Network (ECSN).Source: §1 · p. 2
• Under a positive-semidefinite energy shift and per-bond factorization, the construction contains Slater–Koster tight binding as a special case.Source: §1 · p. 2
• O(3)-steerable stalks and restriction maps make the sheaf Laplacian E(3)- and permutation-equivariant.Source: §1 · p. 2
• Zeroth sheaf cohomology counts non-bonding orbitals and, for alternant systems, is lower bounded by sublattice imbalance.Source: §1 · p. 2
• ECSN strictly generalizes E(3)-equivariant message-passing and cellular networks while inheriting anti-oversmoothing from non-trivial sheaf diffusion.Source: §1 · p. 2
• Numerical validation finds machine-precision embedding and equivariance, correct cohomological counts across eleven molecules, and improved accuracy and rotation robustness over a coordinate baseline.Source: §1 · p. 2
• The paper’s novelty is the sheaf-theoretic formalization and its invariants, rather than equivariant Hamiltonian prediction itself.Source: §1 · p. 2

• Equivariant chemistry networks predict invariant or covariant atomic-graph targets, but do not use higher cells or sheaf structure.Source: §2 · p. 2
• PhiSNet, DeepH, and QHNet predict localized Hamiltonian matrices equivariantly; this paper newly reinterprets the predicted operator as a sheaf Laplacian.Source: §2 · p. 3
• Existing sheaf and cellular-complex methods are not O(3)-equivariant and are not applied to electronic structure.Source: §2 · p. 3
• Related molecular topology work uses density critical-point complexes or persistence-based descriptors, whereas this paper uses chemical cells and sheaf cohomology.Source: §2 · p. 3

3.1 Regular cell complexes from molecules

• Atoms form 0-cells, cutoff pairs form unordered 1-cells, and chosen cycle-basis faces form 2-cells with oriented signed incidences.Source: §3.1 · p. 3

3.2 Cellular sheaves and the sheaf Laplacian

• A cellular sheaf assigns a finite-dimensional inner-product stalk to every cell and a linear restriction map to every incidence.Source: §3.2 · p. 3
• The degree-k Hodge–sheaf Laplacian is Lk = (δk)ᵀδk + δk−1(δk−1)ᵀ, with LF = L0 = (δ0)ᵀδ0 for vertices.Source: §3.2 · p. 3
• LF is symmetric positive semidefinite, and its kernel equals H^0(X; F), the space of global sections.Source: §3.2 · p. 3

3.3 O(3)-equivariant sheaves

• O(3) irreducible stalks model atomic orbitals by angular momentum, while direction-dependent steerable maps are built from Clebsch–Gordan contractions of spherical harmonics.Source: §3.3 · pp. 3–4
• The steerable kernels use learnable weights and radial scalars depending on bond length.Source: §3.3 · p. 4

Hamiltonian–sheaf correspondence

• A localized Hamiltonian with sparse bond-localized blocks becomes a sheaf Laplacian after a PSD energy shift.The construction may use augmented stalks and satisfies LF = H − ErefI.Source: §4.1 · p. 4
• Per-bond SVD factorization matches hopping blocks, while augmented self-incidences match residual on-site terms without violating positive semidefiniteness.The residual construction yields LF = H − ErefI under the PSD assumption.Source: §4.1 · p. 4

Equivariance

• O(3)-steerable restriction maps make the sheaf Laplacian E(3)- and permutation-equivariant, and equivariance passes to diffusion layers and invariant or covariant readouts.Translations act trivially on stalks, while rotations, reflections, and atom relabelings conjugate the operator appropriately.Source: §4.2 · pp. 4–5

Sheaf cohomology

• Choosing Eref at the non-bonding level makes ker LF equal to the eigenspace of H at Eref and makes its dimension deformation-stable.Different reference energies produce sheaves related by stalk-wise orthogonal gauge transformations.Source: §4.1 · p. 4, §4.3 · p. 5
• For alternant systems, dim H0 is lower bounded by the sublattice imbalance, recovering the classical non-bonding π-orbital count.The bound follows from the nullity of the off-diagonal bipartite Hamiltonian.Source: §4.3 · p. 5
• Benzene has dim H0 = 0, whereas cyclobutadiene and trimethylenemethane each have dim H0 = 2.The benzene result is consistent with equal bipartite sublattices and its closed-shell spectrum.Source: §4.3 · p. 5

Higher cells

• Adding ring 2-cells yields L1 = δ0(δ0)⊤ + (δ1)⊤δ1, whose harmonic space H1 captures independent cycles and non-trivial restriction-map closure failures.For the trivial sheaf, dim H1 equals the first Betti number; non-trivial sheaves can signal holonomy or frustration.Source: §4.4 · p. 6

Complex and sheaf construction

• ECSN constructs a regular cell complex from molecular coordinates and species, then assigns atomic stalks and geometry-conditioned steerable restriction maps.The maps depend on invariant features and bond geometry through radial networks.Source: §5 · p. 6

Operator assembly

• The learned restriction maps are assembled into the sheaf Laplacian and optionally normalized as D−1/2LFD−1/2.Here D is the diagonal matrix formed from the vertex blocks of LF.Source: §5 · p. 6

Sheaf diffusion

• Equivariant sheaf diffusion uses gated nonlinearities acting only on O(3) norms, preserving the representation structure during feature propagation.Within-stalk and channel mixing use separate equivariance-preserving transformations.Source: §5 · p. 6

Readout heads

• Invariant readouts predict energies and gaps, covariant readouts output dipoles, forces, or Hamiltonian blocks, and topological readouts report non-bonding and cycle dimensions.The spectral gap is used to estimate the HOMO–LUMO gap, while the Hamiltonian target is recovered as a structured special case.Source: §5 · p. 7

Expressivity hierarchy

• Scalar ECSN diffusion recovers CW-network updates and standard MPNN message passing on the molecular 1-skeleton.These reductions follow by restricting stalks and maps to scalar forms.Source: §6 · p. 7
• Restricting to the 1-skeleton with diagonal steerable blocks recovers an E(3)-equivariant tensor-field MPNN.Source: §6 · p. 7
• ECSN is strictly more expressive because non-trivial sheaves can have H0(X; F) = 0, unlike trivial-sheaf models whose kernels always contain constant sections.Such sheaves can distinguish complexes that trivial-sheaf models cannot.Source: §6 · p. 7

Anti-oversmoothing

• Trivial sheaf diffusion collapses features onto its constant-containing kernel, whereas trivial agreement spaces on cycles prevent this oversmoothing.For non-trivial sheaves, the Dirichlet energy is bounded below on the orthogonal complement of ker LF.Source: §6 · p. 7

Complexity

• With maximum stalk dimension d, assembling and applying LF over T layers costs O(T|E|d2) time and O(|E|d2) memory.An optional spectral readout adds an O(Nd)-dimensional sparse eigenproblem.Source: §6 · p. 7

E1: Hamiltonian embedding

• The scalar-stalk Hamiltonian-to-sheaf reconstruction is exact, with maxmol ∥LF − ˜H∥F = 0 across eleven π-conjugated molecules.For multi-orbital benzene systems, reconstruction error is at most 8.4 × 10−15.Source: §7.1 · p. 8

E2: Cohomology counts

• Computed dim H0 reproduces known non-bonding π-orbital counts across eleven conjugated molecules, including zero, one, and two zero-mode cases.The bipartite lower bound holds for every bipartite system and is tight for odd alternants and trimethylenemethane.Source: §7.2 · p. 8

E3: Equivariance

• Across 200 random rotations, the relative equivariance error is 8.1 × 10−16, and reflection error is exactly 0.Replacing steerable maps with raw-bond-vector maps raises the error to 0.58.Source: §7.3 · p. 9

E4: Learning benefit

• At N = 160, the equivariant model is 58% more accurate on rotated inputs than the coordinate model for the HOMO–LUMO gap target.Its MAE decreases from 0.32 to 0.11 as N grows from 20 to 160, while the coordinate model plateaus near 0.22 and reaches 0.27–0.30 on rotated molecules.Source: §7.4 · p. 9, Table 2 · p. 9

Scope of evaluation

• The evaluation omits large-scale QM9, MD17, and self-consistent Hamiltonian benchmarks, leaving ring-cell H1 benefits and full Hamiltonian-regression data efficiency as predictions.The paper explicitly states these are predictions, not results.Source: §7.5 · p. 9

Reference energy

• A poor reference energy Eref can move the chemically meaningful kernel after the required PSD shift.Choosing Eref at or below the spectrum is required; selecting the non-bonding level makes the kernel chemically meaningful.Source: §8 · p. 9

Gauge freedom

• Restriction maps are identifiable only up to a stalk-wise orthogonal gauge, making the Laplacian better posed for supervision than individual maps.The Laplacian is gauge-invariant, whereas direct supervision on individual restriction maps is complicated.Source: §8 · p. 9

Cycle representation

• Cycle bases for 2-cells are non-unique, so results should use a fixed canonical choice such as the smallest set of smallest rings.Source: §8 · p. 9

Evaluation setting

• The directional-gap comparison uses no rotation augmentation and contrasts the equivariant sheaf model with a coordinate model that degrades on rotated test molecules.Source: Figure 3 · p. 10

Cost and scope

• Sheaf assembly costs d2 relative to scalar GNNs, and genuinely correlated multireference systems remain outside the single-particle model's scope.The stated assumption is single-particle mean-field electronic structure; many-body extensions on higher cells would be required for correlated systems.Source: §8 · pp. 9–10

• The localized-orbital electronic Hamiltonian becomes an E(3)- and permutation-equivariant sheaf Laplacian, with Slater–Koster tight binding as a special case.Source: §9 · p. 10
• Sheaf cohomology supplies topological invariants that count non-bonding orbitals and detect cycle structure in molecular cell complexes.Source: §9 · p. 10
• ECSN strictly generalizes equivariant MPNNs and CW networks while inheriting the anti-oversmoothing guarantee of non-trivial sheaf diffusion.Source: §9 · p. 10
• The framework unifies topological deep learning with equivariant electronic-structure learning and presents sheaf cohomology as a language for chemical-bonding topology.Source: §9 · p. 10
• The contribution is the sheaf-theoretic formalization and its invariants, not equivariant Hamiltonian prediction itself.Source: §9 · p. 10