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

Predicting the electronic Hamiltonian of a molecule directly from its geometry is a central challenge in quantum chemistry and machine learning. Recent equivariant message‑passing networks have shown promise for learning interatomic potentials and molecular properties, while topological deep learning has extended graph networks to cellular sheaves. This paper bridges these two strands by showing that, in a localized atomic‑orbital basis, the single‑particle Hamiltonian (after a simple shift) can be interpreted as the Laplacian of a cellular sheaf built on a regular cell complex derived from the molecular structure.

What You Need to Know

A cellular sheaf assigns a vector space (the stalk) to each cell of a cell complex and linear maps (restriction maps) between stalks of incident cells. In this work the 0‑cells are atoms, the 1‑cells are covalent bonds, and higher‑dimensional cells can encode angles or torsions if desired. The stalk on an atom holds the coefficients of its localized atomic orbitals; the stalk on a bond holds the interaction space between the two atoms’ orbitals.

The molecular Hamiltonian in the chosen basis is assembled from on‑site terms (diagonal blocks) and hopping terms (off‑diagonal blocks) that depend only on the geometry of the bond connecting two atoms. By interpreting the on‑site terms as identity maps on the atomic stalks and the hopping terms as restriction maps on the bond stalks, the full Hamiltonian becomes the sheaf Laplacian: a sum over cells of the stalk identity minus the restriction maps and their transposes. Adding a constant shift to the diagonal makes the operator positive semidefinite, which is convenient for spectral learning.

To enforce equivariance, the restriction maps are taken to be O(3)‑steerable two‑center kernels: functions of the bond vector that transform according to irreducible representations of the rotation group. When these kernels are expanded in low‑order spherical harmonics, they reproduce the classic Slater‑Koster form for σ, π, and δ bonds. Because the kernels depend only on relative positions, the resulting operator is invariant under permutations of identical atoms and equivariant under global rotations and translations (the full E(3) group).

Why It Matters

This interpretation provides a unifying mathematical framework that connects equivariant graph neural networks with tools from algebraic topology. It explains why certain message‑passing architectures naturally capture the physics of tight‑binding models: they are implicitly learning a sheaf Laplacian. Consequently, practitioners can import concepts such as sheaf cohomology, spectral sequences, or localized eigenmodes to analyze and improve their models.

From a practical standpoint, viewing the Hamiltonian as a sheaf Laplacian suggests new ways to impose physical constraints directly into the architecture—e.g., enforcing gauge‑like conditions on restriction maps or incorporating known symmetries of the stalks. It also opens the door to using cohomological invariants as descriptors for transfer learning across chemical space, potentially reducing the amount of quantum‑chemistry data needed to achieve a given accuracy.

Key Details

• Cell complex construction: 0‑cells = atoms, 1‑cells = bonds (optionally 2‑cells = angles, 3‑cells = torsions) based on a distance cutoff.
• Stalks: atomic stalk = space of orbital coefficients on that atom; bond stalk = interaction space spanned by products of the two atoms’ orbital bases.
• Restriction maps: O(3)‑steerable kernels K(r̂) that map the atomic stalk to the bond stalk, where r̂ is the unit bond vector; kernels are expanded in spherical harmonics to guarantee steerability.
• Hamiltonian as sheaf Laplacian: H = Σ₀ (I₀) + Σ₁ (I₁ – R₁ – R₁ᵀ), with I₀, I₁ identity maps on atom and bond stalks, R₁ the restriction map on each bond; a constant shift λI makes H ≥ 0.
• Equivariance proof: Permutation invariance follows from shared kernel parameters across identical bonds; E(3) equivariance follows from the transformation law of steerable kernels under rotations and translations.
• Slater‑Koster connection: When K(r̂) is truncated to l ≤ 2 spherical harmonics, the resulting hopping integrals match the Slater‑Koster formulas for σ, π, and δ bonds.

What’s Next

Future work could extend the sheaf formulation to include multi‑center terms (e.g., three‑cell stalks for angular dependence) and spin‑orbit coupling by enriching the stalks with SU(2) representations. Scaling to larger systems may benefit from sparse sheaf representations and hierarchical pooling inspired by cellular complexes. Additionally, exploring the sheaf cohomology groups of the learned Hamiltonian could yield novel, interpretable descriptors for reactivity spectra or excitation energies, bridging geometric deep learning with traditional quantum‑chemical analysis.

📌 Source: Arxiv Ml