🤖 machine learning

Topological Neural Operators

This paper introduces Topological Neural Operators (TNOs), a principled framework that extends neural operators to cell complexes by leveraging Discrete Exterior Calculus to model cross-dimensional interactions and preserve physical conservation laws, thereby achieving superior accuracy in solving partial differential equations on irregular geometries.

Original authors: Lennart Bastian, Samuel Leventhal, Mustafa Hajij, Tolga Birdal

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lennart Bastian, Samuel Leventhal, Mustafa Hajij, Tolga Birdal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to understand the weather, the flow of electricity, or how a bridge bends under stress. These are all governed by complex physics equations (called Partial Differential Equations, or PDEs).

Traditionally, computer models treat the world like a giant grid of dots. They ask, "What is the temperature at this dot? What is the wind speed at that dot?" They then try to guess how these dots talk to each other.

The Problem:
The authors of this paper argue that this "dot-only" approach misses the most important part of physics: shape and dimension.

In the real world, different physical things live on different shapes:

Temperature lives on points (like a spot on a map).
Wind or electric current flows along lines (like a river in a channel).
Magnetic fields or water flow pass through surfaces (like a net catching rain).
Density fills up volumes (like air inside a balloon).

If you force all these different things to live on just "points," you lose the natural rules that govern them. For example, in physics, the "curl" of a field (how much it swirls) and the "divergence" (how much it spreads out) are mathematically linked. If you squash everything into points, the computer has to re-learn these fundamental laws from scratch, which is hard and often inaccurate.

The Solution: Topological Neural Operators (TNOs)
The authors introduce a new type of AI called Topological Neural Operators (TNOs). Think of TNOs as a computer model that respects the "shape" of the data.

Instead of just a grid of dots, TNOs use a Cell Complex. Imagine a Lego structure:

Some data lives on the corners (vertices).
Some lives on the sticks connecting them (edges).
Some lives on the flat faces of the bricks.
Some lives inside the blocks themselves (volumes).

How it Works (The "Traffic Cop" Analogy):
The paper uses a mathematical framework called Discrete Exterior Calculus (DEC). You can think of DEC as a set of fixed, unchangeable traffic rules for how information moves between these different shapes.

The Fixed Rules (The Traffic Cop): The model knows that information must flow from a point to a line, or from a line to a face, in a specific way. It knows that if you go around a loop, certain things must cancel out (like a river flowing in a circle). These rules are "hard-coded" into the model's architecture. They are the laws of physics built into the machine.
The Learnable Part (The Driver): The AI doesn't learn where the information can go (that's fixed by the shape). Instead, it learns how to transform the data as it travels. It learns how to mix the "wind" data with the "pressure" data to predict the next state.

The "Hierarchical" Upgrade (HTNOs):
The authors also created a version called Hierarchical TNOs (HTNOs).
Imagine you are looking at a map of a city.

The TNO looks at every single street and house (fine detail).
The HTNO creates a "zoomed-out" version of the map (a coarse complex) where neighborhoods are grouped together. It solves the problem on the big picture first, then fills in the details. This helps the AI understand long-distance effects, like how a storm in one part of the city affects the weather in another, much faster.

What They Found:
The team tested this new model on several difficult physics problems:

Fluid Flow: Simulating air flowing over airplane wings (including tricky situations where the air goes supersonic).
Elasticity: Predicting how materials stretch and bend.
Heat and Diffusion: Simulating how heat spreads through irregular shapes.

The Results:

Better Accuracy: TNOs and HTNOs were more accurate than previous state-of-the-art models (like Graph Neural Operators or Fourier Neural Operators) on these irregular shapes.
Physical Sense: Because the model respects the natural "shape" of the data (points, lines, faces), it naturally obeys conservation laws (like energy or mass not disappearing) without needing extra tricks.
Speed: They achieved these results while maintaining the speed advantages of modern AI, often outperforming traditional, slow numerical solvers.

In Summary:
This paper proposes a smarter way to teach AI about physics. Instead of forcing the world into a grid of dots, TNOs let the data live on its natural shapes (dots, lines, surfaces, and volumes). By building the fundamental laws of physics (how these shapes connect) directly into the model's structure, the AI becomes a better, faster, and more reliable simulator for complex real-world systems.

Technical Summary: Topological Neural Operators (TNOs)

1. Problem Statement

Current Neural Operators (NOs), such as Fourier Neural Operators (FNOs) and Graph Neural Operators (GNOs), have achieved significant success in learning solution maps for Partial Differential Equations (PDEs). However, they fundamentally operate on a point-centric abstraction: all physical quantities (potentials, fluxes, densities) are represented as functions defined on nodes or vertices, with interactions modeled via learned kernels or message passing.

This abstraction neglects a critical structural aspect of continuum physics: geometric type. In physical systems, quantities naturally reside on different dimensional supports:

Potentials live on vertices (0-cells).
Circulations/Fluxes live on edges (1-cells).
Fluxes/Vorticity live on faces (2-cells).
Densities live on volumes (3-cells).

Their interactions are governed by differential operators (gradient, curl, divergence) that couple these quantities across dimensions. In compatible numerical discretizations (e.g., Finite Element Exterior Calculus), these relationships are encoded via incidence structures, ensuring conservation laws (e.g., $\text{div}(\text{curl}) = 0$ ) hold by construction. By collapsing all quantities to nodes, existing NOs obscure this structure, limiting their ability to model multi-physics systems, enforce conservation, and capture topological effects like cycles and harmonic modes.

2. Methodology

2.1 Core Framework: Topological Neural Operators (TNOs)

The authors introduce Topological Neural Operators (TNOs), a framework for operator learning on cell complexes (specifically Regular Cell Complexes, RCCs). Instead of representing data solely on points, TNOs represent data as cochains—features defined on cells of varying dimensions (vertices, edges, faces, volumes).

Key Design Principles:

Decoupling Flow and Transformation: The framework separates where information flows (governed by fixed, topological operators) from how it is transformed (learned).
Discrete Exterior Calculus (DEC): Interactions between cochains of different dimensions are modeled using DEC operators:
- Discrete Exterior Derivative ( $d_k$ ): Maps $k$ -cochains to $(k+1)$ -cochains (e.g., gradient, curl).
- Codifferential ( $\delta_k$ ): Maps $k$ -cochains to $(k-1)$ -cochains (metric adjoint of $d$ , e.g., divergence).
- Hodge Laplacian ( $\Delta_k$ ): Combines $d$ and $\delta$ to provide same-rank propagation mediated by adjacent dimensions.
Native Cross-Dimensional Coupling: Information flows explicitly between ranks via incidence-based operators. The architecture learns channel mixing matrices ( $W$ ) and nonlinearities ( $\phi_\theta$ ) applied to these routed features, but the routing topology is fixed by the cell complex.

2.2 Architecture Details

A TNO layer updates hidden cochains $H = (H_0, \dots, H_N)$ on a lifted complex $\tilde{K}$ :
$H_k^{\text{out}} = \phi_\theta \left( d_{k-1} H_{k-1} W^\downarrow_k, \; \Delta^\uparrow_k H_k W^{\Delta,\uparrow}_k, \; \Delta^\downarrow_k H_k W^{\Delta,\downarrow}_k, \; H_k, \; \delta_{k+1} H_{k+1} W^\uparrow_k \right)$
This structure explicitly separates the signal into exact (from $d$ ), coexact (from $\delta$ ), harmonic (from the kernel of $\Delta$ ), and self channels, aligning with the Hodge Decomposition.

2.3 Hierarchical TNOs (HTNOs)

To capture long-range and topology-dependent information, the authors propose Hierarchical TNOs (HTNOs). These incorporate a sequence of coarsened cell complexes ( $K_0 \leftarrow K_1 \leftarrow \dots \leftarrow K_L$ ).

Information propagates across levels via restriction ( $R$ ) and prolongation ( $\Pi$ ) maps.
Ideally, these maps commute with the coboundary operators ( $d$ ), preserving the de Rham structure across scales (analogous to FEEC multigrid).
HTNOs act as learned V-cycles, using fine-level blocks for local cross-degree physics and coarse-level blocks for global topological propagation.

2.4 Theoretical Unification

The paper demonstrates that standard NOs are restricted cases of TNOs:

FNOs are recovered as rank-0 spectral TNOs on uniform grids.
GNOs are recovered as rank-0 graph-quadrature TNOs.
Thus, TNOs provide a unified perspective on operator learning across discretizations.

3. Key Contributions

A Principled Framework: Introduction of TNOs for operator learning on cell complexes, extending neural operators to cochain-valued fields and enabling explicit cross-dimensional modeling via DEC.
Unified Perspective: A theoretical framework that encapsulates existing point- and graph-based NOs as special cases (rank-0 restrictions).
Hierarchical Extension: Development of HTNOs for efficient multi-scale propagation of global and topological information.
Empirical Validation: Comprehensive ablation studies and benchmarks demonstrating that native higher-rank structure and topological inductive biases improve accuracy and physical consistency.

4. Experimental Results

The authors evaluated TNOs and HTNOs on a range of steady-state and time-dependent PDE benchmarks across irregular 2D and 3D meshes.

Steady-State Benchmarks (Poisson-Gauss, Airfoil, Elasticity): TNOs and HTNOs outperformed seven strong baselines (including RIGNO, Geo-FNO, GINO, UPT) on Poisson-Gauss and Elasticity. On Airfoil flow, TNOs were competitive with RIGNO-18.
Compressible Airfoils & Poisson-with-Sines: On datasets spanning subsonic to supersonic regimes (NACA, RAE airfoils), TNOs and HTNOs significantly outperformed RIGNO and GAOT, attributed to the alignment of TNO's Hodge/harmonic biases with the elliptic characteristics of the flow.
Large-Scale Surface PDEs (EmmiWing): On 3D wing surfaces with ~65K query points, HTNO achieved the lowest aggregate L1 error (2.41%), outperforming Transolver, PointNet, and RIGNO, while converging in $\sim6\times$ fewer epochs.
Controlled Studies:
- Anisotropic Darcy: In a setup where face-oriented tensor information is lossy when projected to vertices, TNOs with native rank-2 ingestion (processing face data directly) outperformed vertex-only or projected variants by $\sim0.5\%$ , proving the benefit of respecting geometric support.
- Component Ablation: Removing the harmonic basis or sheaf transport significantly degraded performance on PDEs with non-trivial curl, confirming that these components are synergistic inductive biases rather than additive features.

5. Significance and Claims

The paper claims that TNOs represent a shift from "learning everything" to "learning the transformation of structurally preserved flows." By decoupling the fixed topological flow (incidence, boundary, Hodge operators) from the learnable feature transformation, TNOs:

Respect Geometric Support: They treat physical quantities according to their natural dimension (vertex, edge, face, volume).
Enforce Conservation: They incorporate conservation and compatibility structures (e.g., $\text{div}(\text{curl})=0$ ) into the architecture by construction, rather than relying on data to learn these constraints.
Enable Discretization Transfer: The shared weights across refinements and the use of intrinsic DEC operators allow the model to generalize across different mesh resolutions and geometries.
Unify Operator Learning: They provide a single framework that recovers existing successful NOs as special cases while extending capability to multi-physics and higher-rank systems.

The authors conclude that this approach points toward "scientific foundation models" that can learn operators on structured spaces, unifying geometry, physics, and computation. They acknowledge limitations in scaling to dynamic/adaptive domains and the need for further theoretical analysis of stability and convergence, noting that TNOs are learned approximations and do not guarantee certified physical correctness outside the training distribution.