HodgeNet: Topology-Aware Neural Architectures

• HodgeNet architectures are advanced neural frameworks that incorporate discrete exterior calculus and combinatorial Hodge Laplacians to process data on vertices, edges, and simplices.
• They utilize learnable Hodge stars and spectral methods for adaptive signal propagation and effective filtering across complex geometric structures.
• HodgeNet models excel in diverse applications including flow interpolation, source localization, mesh segmentation, and PDE solving while preserving topological fidelity.

HodgeNet architectures generalize graph neural network (GNN) and mesh-based learning by leveraging the algebraic and topological structure of discrete exterior calculus (DEC), with a central role played by combinatorial Hodge Laplacians. These networks are specialized for data defined not only on nodes or vertices, but also on higher-order structures such as edges, triangles, and higher simplices, enabling modeling of flow signals, spectral geometry, and multi-scale topology-aware feature representations. HodgeNet and its variants have been realized on graphs (Roddenberry et al., 2019), on triangle meshes (Smirnov et al., 2021), within hybrid iterative-neural PDE solvers (Puthli et al., 3 Feb 2026), through spectral attention over simplicial complexes (Huang et al., 2024), and in data-driven operator learning within transformer-like mesh architectures (Nousias et al., 1 Sep 2025). These models share the use of Hodge-theoretic operators—incidence, boundary, and Hodge star matrices—to construct task- or data-adaptive propagation, filtering, or spectral transforms.

1. Discrete Hodge Theory and Operator Construction

All HodgeNet-style architectures build on the DEC formalism: for a simplicial complex $\mathcal{K}$ (graph, mesh), simplices of each order $k$ are enumerated (vertices, edges, triangles, etc.). For each $k$, the boundary operator $B_k$ encodes incidence relations between $k$- and $(k-1)$-simplices, with entries reflecting orientation. The $k$-th combinatorial Hodge Laplacian is

$L_k = B_k^\top B_k + B_{k+1} B_{k+1}^\top.$

For node-based problems, $L_0$ reduces to the standard (combinatorial) Laplacian; for edge-based flow data, $L_1 = B_1^\top B_1$ spans gradient flows and detects cycles (Roddenberry et al., 2019).

On triangle meshes, DEC generalizes the Laplacian to metric-aware, geometry-adaptive form: $k$0 where $k$1 is the combinatorial exterior derivative (gradient) and $k$2 are (possibly learnable) diagonal Hodge-star matrices, encoding volume or dual-area information at vertices and edges, respectively (Smirnov et al., 2021, Nousias et al., 1 Sep 2025).

The design of $k$3 supports:

• spectral decompositions for feature learning,
• propagation of signals across simplices of all dimensions,
• data-driven adaptation of geometric priors via learnable Hodge stars.

2. Core HodgeNet Architectures

Two canonical variants illustrate the HodgeNet approach for edge data:

Hodge RNN (Flow Interpolation on Graphs) (Roddenberry et al., 2019)

The Hodge RNN is designed for edge-signal (flow) imputation/interpolation. Given a partially observed flow $k$4 masked on edges $k$5, the normalized Hodge Laplacian $k$6 (spectral shift operator) is applied recursively:

• At each RNN step: $k$7 (edge-space filtering)
• Nonlinearity and hidden mixing: $k$8, $k$9 The nonlinearity $k$0 is a trainable soft-threshold, preserving flow skew-symmetry.

Hodge AGNN (Source Localization) (Roddenberry et al., 2019)

For classification or source localization, Hodge AGNN builds a sequence of Hodge-shifted flow signals from a learned selection of edge monitors, aggregates by a 1D CNN across shift-depth, and predicts via softmax. This mechanism captures diffusive behaviour and localizes flow sources robustly.

Both architectures emphasize invariance to edge orientation (outputs flip sign if flows are reversed), and directly leverage Laplacian spectra to propagate and aggregate features on the edge space.

3. HodgeNet for Mesh-Based Spectral Geometry

On triangle meshes, HodgeNet generalizes classical mesh descriptors (e.g., HKS, WKS) (Smirnov et al., 2021):

• The input mesh $k$1 is encoded by DEC local operators:
  • $k$2: combinatorial gradient
  • $k$3: learnable diagonal Hodge stars (functions of local vertex features $k$4)
• The Laplacian operator $k$5 is constructed and a truncated eigendecomposition $k$6 is solved for the $k$7 smallest nonzero eigenpairs.
• Per-vertex features are extracted as $k$8, where $k$9 is a learned MLP over spectrum.

End-to-end backpropagation is enabled by closed-form derivatives of losses with respect to Hodge star entries and eigenvectors, which are efficiently approximated over leading spectrum batches. The architecture enables learning of geometry-adaptive operators and descriptors, robustly generalizing across mesh resolutions and classes.

4. Operator Learning and HodgeNet Extensions

Recent models expand HodgeNet concepts beyond fixed-spectrum or edge-only tasks:

• Neural Hodge Corrective Solvers (Puthli et al., 3 Feb 2026): For PDEs on meshes, these hybrid architectures combine standard iterative solvers (Jacobi–Richardson with classical DEC operators) with learned, structure-preserving corrections:
  • DEC-based operators are first fit to data via relative-residual minimization.
  • Neural correction is introduced via small, local CNNs on mesh fields, interleaved in the solver loop.
  • By construction, topological constraints (e.g., exactness of the exterior derivative) are preserved through similarity transforms.
  • The two-stage training process optimizes operator regression first, then convergence of the full hybrid solver.
• Hodge-Laplacian Heterogeneous Graph Attention Networks (HL-HGAT) (Huang et al., 2024): HL-HGAT generalizes GNNs to arbitrary $B_k$0-simplicial complexes, with layers processing signals on $B_k$1-simplices via:
  • Spectral Hodge-Laplacian filters (polynomial transfer functions of $B_k$2 for each $B_k$3)
  • Simplicial projection (mixing signals across dimensions)
  • Simplicial attention pooling (self-attention and cross-attention for resolution coarsening)
  • Full HL-HGAT blocks stack these operators, supporting tasks from edge classification to regression on biological or imaging graphs.
• HodgeFormer (Nousias et al., 1 Sep 2025): This mesh-aware transformer uses multi-head attention to parameterize data-driven Hodge stars $B_k$4 for each $B_k$5, constructing Laplacians $B_k$6 and applying them in parallel stream-wise update on vertex, edge, and face features. The attention is sparsified over mesh neighborhoods to maintain $B_k$7 per-layer cost, sidestepping global eigendecomposition. HodgeFormer achieves state-of-the-art performance on mesh segmentation/classification benchmarks while eliminating preprocessing associated with spectral positional encodings.

5. Backpropagation, Learning, and Computational Aspects

HodgeNet-style models rely on the differentiability of their DEC-based operator stacks and their eigensolvers (Smirnov et al., 2021):

• For spectral descriptors, backpropagation proceeds through the low-order eigenpairs. Derivatives of losses with respect to Hodge-star entries and DEC parameters are implemented via closed, summation-based forms involving the eigenvectors and eigenvalues.
• For edge-based GNNs, the Hodge Laplacian is normalized to have spectral norm $B_k$8 for stability, and trainable nonlinearities ensure invariance and compatibility with edge signal symmetry.
• Models such as HL-HGAT and HodgeFormer avoid explicit (expensive) eigendecomposition during inference or propagation, leveraging either polynomial approximations (as in Chebyshev filters) or localized attention-based learning of operator weights.
• Model size in mesh settings is dominated by the low-rank eigensolve and the few parameter tensors in MLPs ($B_k$9k parameters), and performance scales favorably to large, high-resolution meshes.
• In iterative-neural PDE solvers, the main computational advantage lies in replacing global Newton-style training with local, low-memory correction, maintaining topological fidelity without computing operator Jacobians.

6. Applications and Empirical Performance

HodgeNet and descendants have been evaluated across flow interpolation, source localization, shape classification, segmentation, graph regression, and PDE solving:

• HodgeNet (edge GNN): Robust flow interpolation and community source localization on synthetic graph processes; orientation invariance and accurate completion (Roddenberry et al., 2019).
• Mesh HodgeNet: Spectral features comparable or superior to classical descriptors; strong generalization to out-of-sample mesh classes (Smirnov et al., 2021).
• HL-HGAT: Outperforms classical GNNs on edge-classification, molecular regression, and brain imaging tasks, with improved memory and speed due to attention-based pooling (Huang et al., 2024).
• HodgeFormer: Matches or approaches state-of-the-art mesh classification and segmentation benchmarks (e.g., SHREC-11, COSEG) with significantly reduced preprocessing and runtime (Nousias et al., 1 Sep 2025).
• Neural Hodge Corrective Solvers: Achieve rapid PDE convergence, reducing training and inference time by an order of magnitude compared to Newton-based operator learning (Puthli et al., 3 Feb 2026).

7. Comparative Features and Theoretical Properties

Distinctive properties of HodgeNet-family models include:

• Discrete Hodge Laplacians explicitly model cohomological/topological structure (cycles, cavities, etc.) not accessible to purely node-based message passing.
• Learnable operator and star parameterizations enable direct adaptation to task and data, subsuming both classical (cotangent Laplacian, barycentric areas) and novel, context-sensitive geometries.
• Spectral layers are both interpretable—providing insight into the filtration of signals according to homological classes—and practically tractable via truncated or polynomial approximations.
• In mesh-aware settings, backpropagation through eigen-solvers is efficient and scalable via ARPACK-based routines and summation-based derivatives.

A plausible implication is that future HodgeNet variants may further exploit multi-stream, attention-weighted, or hybrid iterative-neural strategies for complex, topology-sensitive learning on scientific and geometric data.