Topological Neural Operators
Lennart Bastian, Samuel Leventhal, Mustafa Hajij, Tolga Birdal
Abstract
We introduce Topological Neural Operators (TNOs), a principled framework for operator learning on cell complexes that lifts neural operators (NOs) from functions on points and/or edges to topological domains. TNOs represent data as features defined on cells of varying dimension and model their interactions through Discrete Exterior Calculus, enabling explicit cross-dimensional coupling via gradient-, curl-, and divergence-type operators. The key design principle is to decouple where information flows, as governed by fixed topological operators, from how it is transformed (which is learned), yielding models that respect the geometric support of physical quantities and expose conservation and compatibility structure. We further propose Hierarchical TNOs (HTNOs), which incorporate learned coarse complexes to propagate long-range and topology-dependent information. Our framework subsumes existing NOs as a special case, providing a unified perspective on operator learning across discretizations. Across a range of PDE benchmarks, including irregular-geometry flow problems, TNOs and HTNOs improve accuracy; controlled studies further isolate the benefits of native higher-rank and topological structure. Project page: https://circle-group.github.io/research/TNO
AI Impact Assessments
(1 models)Scientific Impact Assessment: Topological Neural Operators
1. Core Contribution
This paper introduces Topological Neural Operators (TNOs), which lift neural operators from point-centric representations to cell complexes of arbitrary dimension. The central insight is that physical quantities have geometric type—potentials on vertices, circulations on edges, fluxes on faces, densities on volumes—and that their interactions are naturally governed by discrete exterior calculus (DEC) operators (gradient, curl, divergence). The key architectural principle is separation of routing from transformation: the DEC operators (d_k, δ_k, Δ_k) prescribe *where* information flows based on incidence structure, while learned channel-mixing matrices determine *how* features are combined. This is formalized through three axioms (cellular intrinsicness, multi-degree coupling, discretization transferability) and realized as topological neural network layers with four channels: exact (d_{k-1}), coexact (δ_{k+1}), Hodge-Laplacian (Δ_k^↑, Δ_k^↓), and self. The paper also introduces Hierarchical TNOs (HTNOs) that add learned coarse complexes for multi-scale propagation, structured as a learned V-cycle for de Rham complexes.
2. Methodological Rigor
The mathematical framework is exceptionally well-developed. The paper carefully builds from regular cell complexes through DEC to the TNO definition, with each step precisely motivated by the physics. Two formal results (Propositions 4.2 and 4.3) prove that FNO and GNO are recovered as rank-0 specializations—establishing TNO as a genuine generalization rather than a parallel framework. The connection to Hiptmair-type Hodge-compatible smoothers and multigrid is articulated carefully, with appropriate caveats that this is a structural analogy rather than a convergence guarantee.
The experimental evaluation spans multiple benchmark suites (RIGNO, GAOT, EmmiWing) covering Poisson, compressible flow, elasticity, and large-scale 3D aerodynamics. Particularly convincing are the controlled ablation studies: (1) the anisotropic Darcy experiment with per-face random tensor orientation is cleverly designed so that vertex projection is *provably* lossy (E[cos(2ϕ)] = 0 under uniform ϕ), isolating a ~0.5pp gain from native rank-2 ingestion on top of ~5pp from the architecture itself; (2) the component ablation on synthetic topologies demonstrates that harmonic basis and sheaf transport are synergistic—sheaf without harmonics actually hurts performance.
One methodological concern: the Hodge decomposition requires computing harmonic projectors P^harm_k, whose cost scales with the number of cells and could be prohibitive for very large meshes. The paper acknowledges this implicitly by noting HTNO subsumes TNO on EmmiWing where per-sample Hodge decompositions on large resampled geometries are impractical.
3. Potential Impact
Immediate impact on operator learning: TNOs provide a principled inductive bias for PDE surrogates that respects the geometric type of physical quantities. The unifying perspective—showing existing NOs as special cases—is valuable for the community's conceptual understanding. The framework naturally handles multi-physics coupling (Maxwell, Darcy, Navier-Stokes) where fields at different ranks interact, which existing methods handle only indirectly.
Broader scientific computing: The connection to finite element exterior calculus (FEEC) and structure-preserving discretizations bridges machine learning with a mature numerical analysis tradition. This could influence how physics-informed ML models are designed for conservation-critical applications (electromagnetics, fluid dynamics, elasticity).
Limitations on immediate practical impact: The overhead of constructing cell complexes, computing Hodge stars, and performing harmonic projections may limit adoption in settings where simple point-cloud or graph methods suffice. The improvements on standard benchmarks, while consistent, are modest in absolute terms (e.g., 1.03% vs. 2.26% on Poisson-Gauss). The most compelling case—native higher-rank ingestion—requires problems where the input data genuinely lives on higher-dimensional cells, which is not the common case in current benchmark ecosystems.
4. Timeliness & Relevance
The paper addresses a genuine gap at the intersection of two active research areas: neural operators for PDEs and topological deep learning. While both fields have grown rapidly, they have remained surprisingly disconnected—TDL architectures have rarely been applied to operator learning, and NOs have not systematically exploited cell-complex structure. The timing is appropriate: the community is moving beyond regular grids to irregular geometries (Geo-FNO, RIGNO, GAOT), and TNOs provide the natural next step by encoding the geometric type of physical quantities directly.
5. Strengths & Limitations
Strengths:
Limitations:
Generated Jun 9, 2026
Comparison History (21)
Paper 1 addresses fundamental efficiency bottlenecks in language modeling—the dominant paradigm in AI—with both theoretical guarantees and practical speedups over FlashAttention 2, a widely-used baseline. Its impact spans long-context prefill, KV cache compression, and decoding, all critical problems at scale. Paper 2 introduces a mathematically elegant framework for operator learning on cell complexes, but targets a narrower audience (PDE/scientific computing). Given the enormous scale of LLM deployment and active research on efficient attention, Paper 1's practical and theoretical contributions are likely to have broader near-term impact.
Topological Neural Operators introduces a fundamentally new mathematical framework that extends neural operators to cell complexes using Discrete Exterior Calculus, providing a principled unification of existing neural operator architectures. This has broad impact across scientific computing, physics-informed ML, and computational geometry. Paper 1, while offering a useful conceptual lens for SFT, is more incremental—reframing existing loss functions as target distribution design and proposing modest improvements. Paper 2's novelty in connecting topology, geometry, and operator learning opens richer research directions with applications across multiple scientific domains.
While Paper 1 offers practical efficiency and performance gains for LLM reasoning, Paper 2 introduces a fundamental, unified mathematical framework (Topological Neural Operators) for scientific machine learning. By subsuming existing neural operators and integrating Discrete Exterior Calculus, Paper 2 provides exceptional methodological rigor and broad applicability to physics and engineering, suggesting a deeper and more lasting scientific impact across multiple disciplines.
Topological Neural Operators introduces a fundamentally new mathematical framework that unifies operator learning on cell complexes using Discrete Exterior Calculus, subsuming existing neural operators as special cases. This has broader impact across scientific computing, physics-informed ML, and computational geometry. The principled decoupling of topology from learned transformations and the respect for conservation/compatibility structures address deep theoretical questions. Paper 1, while practically valuable for compiler optimization, addresses a narrower application domain with incremental improvements to existing auto-scheduling frameworks.
Paper 1 is a foundational review that proposes a unifying framework (REO) and a phase diagram for the rapidly expanding field of data-driven equation discovery. By synthesizing diverse AI-based methodologies and outlining future challenges in revising scientific theories, it has the potential to broadly shape the research agenda across physics and adjacent sciences. While Paper 2 offers a strong technical advancement in neural operators, Paper 1's conceptual synthesis and broad interdisciplinary relevance give it a higher potential for widespread scientific impact.
Paper 1 is more novel and broadly impactful: it generalizes neural operators to cell complexes with principled topological structure (DEC, cross-dimensional coupling) and proposes hierarchical coarse-complex learning, potentially affecting operator learning, scientific ML, PDE solvers, and geometry processing across many domains. It also reports improvements on multiple PDE benchmarks including irregular geometries, suggesting stronger methodological breadth and applicability. Paper 2 is interesting but narrower (force prediction on a single NaCl aqueous testbed) and shows modest MAE gains with limited validation, so its likely impact is more specialized and less certain.
Paper 1 introduces a novel mathematical framework (Topological Neural Operators) that extends neural operators to cell complexes using Discrete Exterior Calculus, with demonstrated empirical improvements on PDE benchmarks. This combines topological/geometric deep learning with operator learning in a principled way, offering both theoretical contributions (subsumption of existing NOs) and practical advances for scientific computing. Paper 2 provides a valuable conceptual framework for evaluating foundation model research validity, but is primarily a methodological commentary rather than introducing new computational methods. Paper 1's technical novelty and direct applicability to physics-based problems give it higher potential for lasting scientific impact.
Paper 2 likely has higher impact: it introduces a broadly applicable, principled framework (Topological Neural Operators) that generalizes neural operators to cell complexes via discrete exterior calculus, enabling cross-dimensional coupling and better physical structure (conservation/compatibility). This is timely for scientific ML and PDE surrogate modeling, with clear real-world applications in engineering/physics on irregular geometries and meshes, and breadth across computational physics, geometry processing, and ML. Paper 1 is innovative for transformer memory/metaplasticity but is narrower and more model-specific, with less immediate cross-domain applicability.
Paper 1 introduces a principled mathematical framework (Topological Neural Operators) that generalizes neural operators to cell complexes using Discrete Exterior Calculus, providing a unified theoretical perspective that subsumes existing approaches. It addresses fundamental challenges in operator learning for PDEs with broad applicability across scientific computing. Paper 2 addresses a narrower engineering problem (cold-start forecasting in cloud-edge systems) with incremental contributions (data mixing, a dataset). Paper 1's novelty, theoretical depth, methodological rigor, and breadth of potential impact across computational physics and machine learning give it substantially higher scientific impact potential.
Paper 2 likely has higher scientific impact: it introduces a broadly applicable framework (Topological Neural Operators) that generalizes neural operators to cell complexes with principled DEC-based cross-dimensional coupling, unifying multiple discretizations and directly targeting PDE/operator learning—an active, high-impact area spanning ML, scientific computing, physics, and engineering. Its methodological framing (fixed topological operators + learned transforms) and hierarchical extension can influence many downstream models and domains. Paper 1 is novel and useful for constraint-satisfying LLM generation in geometry, but its impact is more niche and benchmark-specific.