acceptodds
Under review as a conference paper at ICLR 2027

Topological Neural Operators

64 views

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 discrete 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 coarse complexes to propagate long-range and topology-dependent information. Our framework recovers standard point- and graph-based NOs as restricted cases, providing a unified perspective on operator learning across discretizations. Across a range of PDE benchmarks, including problems on irregular geometries, TNOs and HTNOs improve accuracy and physical residuals. Controlled studies further isolate the benefits of incorporating native higher-rank and topological structure.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Neural Operators, Physics Guided Machine Learning, Topological Deep Learning, Geometric Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_4GmYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces topological neural operators. The proposed work aims to generalizes neural operators, which operate functions on points, to operators on cell complexes. This operator is constrained to respect the signal topology through DEC, and retains flexibility through a learned metric.

Critical strengths
  • The problem considered in this paper is of significant interest to the learning community
  • The approach offered is principled, and effective in several regimes
  • The paper is easy to read in most places
Critical weaknesses
  • There doesn’t appear to be a dedicated related work section which makes it difficult to assess how this work sits in relation to existing literature, and the baselines presented in the empirical analysis.
  • Related papers [1-2] are not compared against, or discussed (as I understand ICLR does not consider ICML publications contemporaneous)
  • It is not clear to me how to understand the empirical analysis beyond performance, since the baselines are not discussed in a related work section and there is no discussion of the interpretation of the results.

[1] Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs. Shaffer, B.D., Koohy, S., Kinch, B., Hsieh, M.A. & Trask, N.. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 2026 [2] Topology-Preserving Neural Operator Learning via Hodge Decomposition, Dongzhe Zheng, Tao Zhong, Christine Allen-Blanchette Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research, 2026

Questions for authors
  • Should H_out be positive definite?
  • Does the learned nonlinear operator preserve the structure of H_out?
  • Does the fact that the raising and lowering transfer maps don’t commute impact whether TNO respects (P1-3)?
3 comments

We thank the reviewer for finding the approach principled and effective in several regimes. We address the positioning, interpretation, and structural properties below.

W1. Related-work organization.

Can a dedicated related-work section position the method and baselines?

We agree that placing the extensive related-work discussion in the appendix can make the contribution and baseline choices harder to assess. We have now also added a related work section in the main text that distinguishes neural operators on irregular geometries/topologies, higher-order networks, and methods that explicitly enforce differential constraints. This section now explains both the relationship to prior work and the baselines we evaluate against, and points to the extended related works in the appendix.

W2. Additional comparisons.

How does TNO compare with Shaffer et al.'s structure-preserving learning and Zheng et al.'s Hodge-based operator?

We thank the reviewer for highlighting these closely related works. Our submission includes a direct comparison with Zheng et al.’s Hodge-based operator (HSD) in Table 6, covering both configurations for the published network width and parameter-equalized variants, which are evaluated across all seven 2D benchmarks. Appendix E.3 details the implementation, tuning, and evaluation protocol.

We further thank the reviewer for pointing us to Geo-NeWF. It shares our motivation to exploit compatible differential structures, but learns geometry-conditioned reduced finite-element spaces and constitutive laws rather than learning a solution operator via cochain updates, as in TNO. We have now included Geo-NeWF as relevant related work, but prioritize HSD and HOGNN as baselines as they more directly compare to our TNO instantiation's (Eq 9) Hodge-based operator design and DEC-based architecture, respectively (Table 6).

W3. Interpretation of results.

What do the results establish beyond predictive accuracy?

Beyond improvements in predictive accuracy on public benchmarks, our central contribution is a way to formulate operator learning for fields that inherently live at different ranks of a complex. Potentials, circulations, and fluxes naturally live on vertices, edges, and faces, with interactions defined through discrete differential operators. This is established in the DEC literature, and we formalize it here in the language of neural operators. Topological neural networks (TNNs) provide architectures for learning representations on cell complexes. TNOs instead formulate learning operators between cochain-valued physical fields and use TNN layers as their computational building blocks. Our approach separates the geometric relationships between fields from the learned feature transformations. The same parameters can then be applied to different meshes by assembling the geometric operators for each mesh. The formulation also connects these constructions to existing neural operators, with Fourier and graph neural operators recovered as special cases. Our contribution is therefore the integration of DEC operators and topological message passing into a unified framework for neural operator learning on cell complexes, rather than either component individually. We summarize these contributions in our response to hvUN, W4.

Sections 5.3–5.4 in the submitted manuscript already examine why these choices matter. Our Oriented-Darcy experiments demonstrate how TNOs can natively represent multi-rank input information, while the component ablations examine how local cross-rank interactions and global spectral communication contribute to predictive performance. Additional cellular message-passing ablations now also justify the benefit of DEC routing itself (gmQ3, W2 / Q2 (a)), and a cross-rank ablation examines the contribution of interactions between ranks (gmQ3, W2 / Q2 (b)).

We have now made this interpretation clearer in the paper. We do not claim that learned model updates preserve conservation laws, as discussed in Q2 below.

Q1. Positive definiteness.

Should HoutH_{\mathrm{out}} be positive definite?

In Eq. 8, Houtk∈Rnk×dhH_{\mathrm{out}}^k\in\mathbb{R}^{n_k\times d_h} contains the updated features on each of the kk-cells. Positive definiteness is therefore not a condition on this block feature update. The positive-definite objects in the DEC construction are the Hodge stars MkM_k, which define the cochain inner products, yielding positive semidefinite Hodge Laplacians in these inner products.

If the question concerns the assembled block operator in Sec. 4.2, positive semidefiniteness of its diagonal Laplacian blocks does not imply positive definiteness of the coupled operator. We do not require this property. We use the block operator to prescribe geometric routes for feature transport, drawing analogies to how DEC couples dimensioned physical quantities through boundary operations.

Q2. Structure under nonlinear updates.

Does the learned nonlinear operator preserve HoutH_{\mathrm{out}}'s structure?

The layer preserves the assignment of features to cochain degrees and their interactions in the language of DEC. For example, dk−1Hk−1d^{k-1}H^{k-1}, δk+1Hk+1\delta^{k+1}H^{k+1} and ΔkHk\Delta_kH^k are all kk-cochain-valued quantities. We combine these quantities at each kk-cell using the same learned feature transformation across cells of that degree (see Eq 9).

A generic cellwise nonlinearity does not, however, preserve closed, exact or harmonic subspaces. Even if dkhk=0d^kh^k=0, we don't necessarily have dkϕ(hk)=0d^k\phi(h^k)=0. Similarly, the identities of the fixed DEC operators do not impose the corresponding identities on independently decoded outputs. An additional ablation now illustrates this latter distinction by constructing outputs using dd and δ\delta, which enforce their differential relations but increase their prediction errors (Reviewer 9Qbx, W1 / Q2 (b)).

Q3. Noncommuting transfer maps.

How does noncommutation of raising/lowering maps affect P1–P3?

Noncommutation does not affect P1–P3. We interpret this as two separate questions of compatibility.

Within a complex, the raising and lowering maps dd and δ\delta do not need to commute (dδd\delta and δd\delta d are the two contributions to the Hodge Laplacian), and P1–P3 do not require them to be equal. For the fixed DEC operators, the relevant identities are dk+1dk=0d^{k+1}d^k=0 and the Hodge-adjoint relation between dd and δ\delta.

Between resolutions, restriction maps RkR^k would additionally need to satisfy dckRk=Rk+1dfkd_c^kR^k=R^{k+1}d_f^k to be compatible in the de Rham sense, and likewise for prolongation. Our HTNOs build the hierarchy on general combinatorial complexes, which do not guarantee this de Rham structure, so we do not enforce this condition. P1–P3 are relabeling equivariance, incidence-supported coupling, and weight sharing, and each HTNO level satisfies them. Without the commuting condition, the hierarchy does not guarantee that exact cochains remain exact after transfer, so the coarse levels act as a learned coarse correction, as we discuss in Appendix D.1.

We thank the reviewer for their questions and are happy to follow up. Given the additional evaluations and exposition, we would appreciate it if the reviewer would consider raising their final score once their concerns have been addressed.

Dear Reviewer 4GmY,

As the discussion period closes on November 18 (AoE), we would be grateful to hear whether the new related-work section in the main text (W1), the comparison with Zheng et al.'s HSD in Table 6, the positioning of Geo-NeWF (W2), and our answers on positive definiteness, nonlinear updates, and noncommuting transfer maps (Q1–Q3) address your concerns. We appreciate the time you spent reviewing our work and are happy to clarify any remaining points in the time we have left.

Best, The authors

Reviewer_hvUNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Topological Neural Operators (TNOs). These store inputs, hidden states and outputs on the vertices, edges, faces and volumes of a mesh, rather than on points only. Each layer moves features between and within these ranks through fixed operators from discrete exterior calculus (DEC): signed incidence matrices weighted by mesh geometry. The network learns the channel mixing along each route.

Optional components are:

  • a projection onto the harmonic subspace of each Hodge Laplacian;
  • learned transport along incidences (the copresheaf variant);
  • a k-means coarse level arranged as a V-cycle (HTNO).

FNO and GNO are shown to be special cases that operate on vertices only.

On steady-state benchmarks, TNO or HTNO is best on six of seven public 2D tasks and on the EmmiWing aggregate. Two synthetic studies show large gains over a vertex-based message-passing network (MPNN): one has inputs stored on faces (Oriented-Darcy), the other has outputs on several ranks (curl–curl).

Critical strengths
  • A sound design for operators on meshes. Fluxes and circulations are stored on faces and edges and coupled through the discrete gradient, curl and divergence. Signed incidence matrices handle edge and face orientation consistently.
  • Global communication through harmonic projection. A good design to bring large-scale dynamics into the neural operator
  • Transfer across meshes. The weights do not depend on the mesh, and the harmonic information enters as a projection, so the model transfers across meshes. HSD (Zheng et al., ICML 2026) ties weights to eigenmode indices and fails on the same variable-mesh tasks.
  • Strong results on public benchmarks. TNO and HTNO roughly halve the error of RIGNO-18 (Mousavi et al., NeurIPS 2025) on Poisson-Gauss and Elasticity. On Oriented-Darcy, feeding the face field at its native rank clearly helps.
  • The appendix examines physical residuals, cross-rank consistency and resolution transfer. It also compares against related topological methods, including HOGNN (Liao et al., NeurIPS 2025).
Critical weaknesses
  1. Time-dependent PDEs are not tested.

    • Section 3 formulates evolution problems, and Maxwell and wave equations motivate the design, but all experiments are steady-state. Only the steady subsets of the RIGNO and GAOT suites are used, although both suites contain time-dependent problems. Without experiments on time-dependent systems, it is difficult to assess the significance of the proposed architecture for the physical problems that motivate it. The claims about its applicability to such systems should therefore be qualified.
    • Steady elliptic problems favour global communication, which the harmonic channel and the V-cycle provide. Advection-dominated problems would test local transport and how error grows over a rollout.
    • Inference and rollout costs are not reported. On moving meshes, the per-mesh eigendecomposition would add further cost.
  2. The ablations do not separate the DEC design from simpler alternatives.

    • The baseline is an MPNN on vertices only, with no edge or face nodes.
    • Without the harmonic channel, TNO performs close to this baseline, and replacing fixed DEC transport with learned transport changes little. Most of the gain comes from the harmonic channel.
    • Since the Darcy target has no harmonic component, that channel may act mainly as a global communication route.
    • A fair control would be a message-passing network on the same vertex, edge and face cells, as in CW Networks (Bodnar et al., NeurIPS 2021), with the same global route and hierarchy at matched cost.
  3. The multi-rank claim rests on synthetic linear problems.

    • On the public benchmarks, inputs and outputs live on vertices; edges and faces carry only hidden features.
    • Native higher-rank data appear only in Oriented-Darcy and curl–curl, two linear problems generated by the authors.
    • A realistic coupled-field problem, such as time-domain Maxwell or mixed-form Darcy, would show whether the multi-rank design helps in practice.
  4. The novelty relative to topological networks is overstated.

    • The building blocks already exist: multi-rank message passing (CW Networks), separate lower and upper Hodge-Laplacian channels with cross-rank terms (SCCNN; Yang, Leus & Isufi, TMLR 2025), DEC-based operator learning (HOGNN).
    • The new step is combining them into a mesh-transferable operator with a harmonic channel, and testing it on current operator benchmarks.
  5. The baseline comparisons are hard to interpret.

    • Most baseline numbers are copied from the RIGNO paper, while TNO is tuned per dataset.
    • GAOT's public code does not reproduce its published results, as the authors document.
    • The main tables report no seed variation.
  6. Two statements do not match what is implemented.

    • Harmonic projector is defined as a projection onto the null space of the Hodge Laplacian, but the benchmarks use eight modes on domains with zero or one hole. If non-zero modes are kept, the channel is a global spectral route and not only a topological one.
    • The introduction says TNOs "incorporate conservation and compatibility structure into the architecture". Appendix H states that predictions are not constrained to conserve anything.

The paper is hard to follow for a general machine-learning reader.

  • The layer itself is simple: message passing between vertex, edge and face nodes with signed, geometry-weighted operators, plus a global projection. You have to get through cochain spaces, the de Rham complex, copresheaves and a "rigidity spectrum" to read the paper. The main idea could have been introduced using simpler languages.
  • 49-page appendix is daunting
  • A plain description of one layer update on a small mesh, early in the main text, would widen the audience.

I would consider raising my score if the authors can provide time dependent nonlinear PDEs and ablation tests to show the value of their designs upon rebuttal.

========

  • Bodnar et al. Weisfeiler and Lehman go cellular: CW Networks. NeurIPS 2021.
  • Liao, Wu & Li. Boundary-value PDEs meet higher-order differential topology-aware GNNs (HOGNN). NeurIPS 2025.
  • Mousavi et al. RIGNO: A graph-based framework for robust and accurate operator learning for PDEs on arbitrary domains. NeurIPS 2025.
  • Wen et al. Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains. NeurIPS 2025.
  • Yang, Leus & Isufi. Hodge-aware convolutional learning on simplicial complexes (SCCNN). TMLR 2025.
  • Zheng, Zhong & Allen-Blanchette. Topology-preserving neural operator learning via Hodge decomposition (HSD). ICML 2026.
Questions for authors
  1. Do the eight harmonic modes contain only null-space modes, or also low-frequency non-zero modes?
  2. What are the inference and rollout costs relative to the MPNN?
6 comments

We thank the reviewer for finding the mesh-operator design sound and for recognizing the public benchmark results. We address the temporal scope, component attribution, and presentation below.

W4. Novelty and contribution.

What is new relative to CW Networks, SCCNN, and HOGNN?

Let us first clarify the scope of our contributions. We present a new operator-learning framework between geometrically typed cochain spaces:

  1. Definition 4.1 (P1–P3) defines maps between geometrically typed cochain spaces. The incidence structure of each complex fixes the support of information routing. The model learns channel mixing and, in the copresheaf variant, the maps along incidences, with weights shared across discretizations.
  2. We provide a unifying view of neural operators. Fourier and graph neural operators arise as rank-0 special cases (Props. 4.2–4.3), which places TNOs within operator learning rather than graph learning. Earlier unifications (Kovachki et al., 2023) relate neural operators for vertex fields and have no notion of a field on an edge or a face.
  3. The spectral channel applies a single learned filter to all retained modes, so no weight is tied to a mode index and the same parameters apply to meshes with different eigenbases. HSD (Zheng et al., 2026) directly links weights to mode indices.
  4. Table 6 compares to conceptually related architectures on the seven public 2D benchmarks. Our proposed TNO significantly outperforms existing conceptually related works, and we analyze why in App. E.3.
  5. TNOs predict fields on several cell ranks at once on unstructured meshes of varying geometry and topology (see our answer to W3).

To our knowledge, no prior work incorporates the following properties: fields on cells of several ranks that are coupled through DEC operations, a learned solution operator with weights shared across discretizations, and sufficient generality to encompass Fourier and graph neural operators as rank-0 special cases. We agree that the individual DEC and message-passing ingredients are established. CW Networks and SCCNN learn representations on cell and simplicial complexes for molecular property prediction, simplex prediction, and trajectory prediction, but are not generally neural operators. Our architecture couples cochain degrees through exterior derivatives and codifferentials, separates same-degree propagation into upper and lower Hodge-Laplacian channels, and incorporates a global spectral channel built from each mesh's Hodge-Laplacian eigenmodes. The spectral channel applies a single learned map across modes, whereas HTNO additionally propagates information through a hierarchy of coarsened complexes. These global and multilevel transformations are also absent from DEC-HOGNN's reported architecture. We compare its official implementation against TNO and HTNO on seven public 2D benchmarks, including datasets with per-sample meshes (App. E.3, Table 6).

W1 (a–b). Time-dependent scope and temporal evaluation.

What supports the time-dependent claims when all experiments are steady-state? Can nonlinear, advection-dominated temporal PDEs test local transport and rollout error?

We agree that our experiments do not test time-dependent problems. Every result in the paper concerns steady-state solution operators, evaluated on public benchmarks with 1K to 65K nodes and on three controlled synthetic studies (Section 5). However, the framework itself is not tied to steady state. Definition 4.1 specifies maps between cochain spaces on a complex, and a next-step map ut↦ut+1u_t\mapsto u_{t+1} is such a map, so a TNO applies to it unchanged as an autoregressive model.

We see this extension as very promising but deliberately restrict the scope of the present work to the steady-state setting, due to the already extensive exposition. Extending to nonlinear PDEs and autoregressive rollout presents numerous design decisions and model ablations. The nonlinear, advection-dominated problems the reviewer suggests are the right benchmarks for it, and we plan a dedicated follow-up on temporal TNOs. We kindly note that the limitations of the submitted manuscript already list scaling to dynamic and adaptive domains as future work, but we have now clarified this further by explicitly stating the steady-state scope in the conclusion and limitations.

W1 (c) / Q2. Inference and rollout costs.

What are inference and rollout costs versus MPNN, including spectral recomputation on moving meshes?

The submitted manuscript already reports spectral preprocessing and training-step costs in Appendix G, “Computational cost” (please see the computational-cost table). These measurements compare TNO variants with an MPNN at approximately 5.36M parameters on Darcy. They quantify preprocessing and training overhead for the steady-state models.

Following the reviewers' suggestions, we additionally measured warmed device-forward passes on curl–curl for the DEC and generic cellular models from gmQ3, W2 / Q2 (a) (9.77M parameters each without the spectral block, 10.44M and 10.48M with it). On a single GH200, median times over 150 synchronized calls, averaged over three seeds, were 2.86/6.55 ms per batch of two for DEC/generic routing without the spectral block and 5.29/7.35 ms with it. Under the same protocol, the MPNN of Table 3 (right, 10.37M parameters) takes 2.21 ms. Generic routing evaluates a learned message for every incidence, whereas DEC routing sums over incidences with fixed signs and applies its learned map once per cell. These timings exclude setup/JIT compilation, data preparation, physical postprocessing, and eigenbasis construction.

On a fixed mesh, the basis can be reused. Moving meshes, however, would require updating geometry-dependent operators and, for spectral TNO, recomputing the basis when these change. As stated in W1 (a–b), we consider temporal models and thus moving-mesh rollouts outside the scope of this work, but have now additionally mentioned this in the limitations. HTNO avoids the spectral decomposition altogether by using hierarchical long-range communication, and is thus more suitable for large-scale datasets such as EmmiWing, which we assess in Table 2b.

With the spectral block enabled, DEC routing also reduces edge/face relative L1 (please see gmQ3, W2 / Q2 (a)), so in this steady-state comparison, DEC routing gives lower prediction errors while having a faster forward pass. We have included these timings in Appendix G ("Computational cost").

W2 (a). Comparisons to cellular baseline.

Does TNO outperform cellular message passing with identical inputs, global communication, hierarchy, and cost?

We now also compare TNO with a generic cellular message-passing model using the same inputs and training protocol (Table 3b, with the model definition in Appendix F.7). DEC routing reduces relative L1 across three seeds (gmQ3, W2 / Q2 (a)), with a faster forward pass (W1 (c) / Q2).

W2 (b). Global communication and topology.

Are the gains mainly global communication rather than topology, especially when Darcy's target has no harmonic component?

We observe that global information exchange contributes substantially to network performance. In the cellular comparison mentioned above, adding the spectral channel reduces errors for both generic cellular MP and DEC transport. However, providing both models with the same spectral channel still leaves DEC routing with an accuracy advantage, so access to global information alone cannot explain the difference.

On Darcy, where the target has no harmonic component, the gain from the spectral channel (5.6 pp in the component ablation of Section 5) comes from global low-frequency communication. The channel uses 16 modes, most with positive eigenvalues, and positive modes alone outperform the kernel alone on both Darcy and curl–curl (gmQ3 W1 / Q1).

We also isolate cross-rank interactions while keeping the same diagonal copresheaf maps and parameter count. Disabling cross-rank transport substantially increases the curl–curl edge relative L1 (please see gmQ3, W2 / Q2 (b)). Together, these controls support the contributions of both global spectral communication and structured interactions between cochain degrees.

W3. Realistic coupled fields.

Does native multi-rank learning help on a realistic coupled problem beyond synthetic linear PDEs?

Our public benchmarks evaluate prediction accuracy on problems including hyperelastic deformation and compressible flow, while our synthetic multi-rank problems let us isolate the contribution of interactions between cochain degrees. The ablations in W2 (a–b) further support this contribution on curl–curl.

As a further multi-field example, we extend Oriented-Darcy (Section 5, Table 3a) to joint prediction of pressure on vertices and normal flux on edges. We keep the dataset, split and training protocol identical. This splits the solution operator into one component for each physical quantity, as defined on their native cells. This also presents an interesting connection to Operator-splitting approaches, which compose such separately learned components (Serrano et al., 2026), making them useful independently of accuracy gains. We have now cited these.

TNO, native rank-2 inputPressure L1 (%)Flux L1 (%)
Pressure only (Table 3a, reference)1.68 ± 0.07–
Pressure and flux (new)1.59±0.10\mathbf{1.59 \pm 0.10}5.37±0.15\mathbf{5.37 \pm 0.15}

Adding the flux output does not degrade the pressure prediction (1.59 ± 0.10% against 1.68 ± 0.07%).

Graph-based message-passing networks have no face fields, and lack oriented cochains on edges, so they cannot natively predict fields on vertices, edges and faces together as a TNO does in these examples. Our MPNN baseline (Table 3b) evolves features on vertices only, and requires additional prediction heads for edge and face values. While there are benchmarks with targets on several cell types on regular grids, (e.g., the staggered Maxwell and Navier–Stokes data of PDEArena), we are not aware of these for irregular-geometries. HSD (Zheng et al., 2026) and DEC-HOGNN (Liao et al., 2025), for instance, also only predict vertex targets. So to our knowledge, TNOs are the first neural operator class evaluated on simultaneous prediction across cell ranks on unstructured meshes of varying geometry and topology, and we see this as an important path forward for many problem settings in the field.

W5 (a). Benchmark protocols.

Are published RIGNO and reconstructed GAOT comparisons on comparable protocols?

For the RIGNO benchmarks, we use the suite's original splits and standardization. GAOT's public release does not provide the exact split indices, so we reran the relevant methods on shared reconstructed splits, using the same graph construction and validation-based checkpoint selection. We explicitly describe this as a comparison based on what is publicly available (see App. F), rather than reproduction of the published GAOT table.

W5 (b). Tuning budgets.

Is tuning effort comparable between TNO and the quoted baselines?

We deliberately evaluate on public benchmarks, so that most baseline numbers are those the original authors tuned and reported themselves. When we re-ran a method (marked † in Tables 1–2, for example where public splits were not available), we used the authors' official implementation and published optimizer settings, and tuned the epoch budget and the per-dataset width on the validation split (App. E.3). This was the case for RIGNO and GAOT, the two strongest related neural operators for large unstructured meshes, as well as Transolver. For the topological baselines, we tuned the hidden width of HOGNN per dataset, and trained SNN-PDE and HSD both at the widths used by their authors and at a parameter count comparable to ours (Table 6, App. E.3). We conducted our model development primarily on the synthetic benchmarks and those in Table 1, but did not tune extensively on the other datasets aside from depth, width, learning rate, neighborhood size, and graph construction (App. F). For all models, including ours, we select the checkpoint with the lowest validation relative L1 error and report its error on the test set.

Reviewer_9QbxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Topological Neural Operators (TNOs), representing physical quantities as cochains on vertices, edges, and faces, with information routed through DEC exterior derivatives, codifferentials, and Hodge Laplacians. The model includes harmonic projection and learnable copresheaf transports. HTNO constructs a coarse level through vertex clustering to handle long-range interactions. Experiments cover public Poisson, hyperelasticity, and compressible-flow benchmarks, together with synthetic Darcy, Oriented-Darcy, and Curl–curl tasks. The method achieves low prediction errors on several public tasks, while harmonic projection produces the largest improvements in the synthetic studies.

Critical strengths
  1. Vertex-, edge-, and face-supported physical quantities have different geometric meanings. Connecting them through exterior derivatives and codifferentials gives the network's propagation paths a physical interpretation and provides a natural representation for mixed physical systems.

  2. The public-benchmark results are strong. HTNO obtains 1.70% on Elasticity versus GAOT's 4.07%, 0.52% on PCS versus 0.72%, and 2.41% on EmmiWing versus Transolver's 2.55% (Tables 1–2). The parameter-matched ablations, spectral-input controls, residual measurements, and cost analysis also provide useful diagnostics.

Critical weaknesses
  1. The roles of the main components remain unclear. Darcy error falls from 8.59% for the parameter-matched MPNN to 7.86% for basic TNO, then to 2.29% with the harmonic component. Curl–curl edge error likewise falls from 18.47% to 2.87% (Table 3b). For Curl–curl, the stated discrete equation already has an analytic solution on the exact harmonic subspace, but there is no comparison with a baseline that handles this part explicitly and learns the remainder. Oriented-Darcy also gives TNO direct face inputs while averaging them onto edges for the MPNN, leaving the role of DEC propagation unresolved. The public tasks lack component ablations within a common backbone, and TNO and HTNO differ in depth and other settings (Appendix F.1). Given the connections to existing DEC-based higher-order networks and Hodge-based neural operators, these controls are important for establishing TNO's specific contribution.

  2. HTNO's coarsening strategy lacks direct validation on native multi-degree tasks. Tables 12–13 evaluate TNO variants but omit HTNO. On Darcy, TNO with harmonic projection and copresheaf transport degrades from 2.13% at 1,000 vertices to 26.08% at 4,000, leaving the benefit of the hierarchy unanswered. In Appendix D.1, only vertex states are explicitly restricted and prolonged; coarse edge and face features are regenerated from coarse vertex states, rather than obtained through compatible transfers of independent higher-degree cochains. There is no quantitative evaluation of how much circulation and flux information this retains. Each synthetic dataset contains only 100 meshes of approximately 500–1,000 vertices (Table 8). Aggregate accuracy on larger public tasks does not establish whether edge and face information survives coarsening.

Questions for authors
  1. Which cochain degrees, harmonic/spectral channels, copresheaf maps, and coarse levels are active in each model in Tables 1–2? What are the component ablations on a representative public task with the backbone and budget held fixed? Does “r=8r=8 harmonic modes” in Appendix F.1 mean the zero eigenspace or a truncated low-frequency subspace, and what boundary conditions and truncation rule are used?

  2. For Curl–curl's (Δ1+εI)A=b(\Delta_1+\varepsilon I)A=b, the exact harmonic projector PHP_H of the same boundary-treated operator gives PHA=ε−1PHbP_HA=\varepsilon^{-1}P_Hb. Do the labels follow this relation or treat the harmonic part separately? How much advantage remains against an analytic harmonic channel consistent with label generation, with the learned component restricted to the complementary subspace? A second direct control is to predict only AA and derive δ1A,d1A\delta^1A,d^1A: how do field errors and system residuals compare? The 499% residual in Table 11 may reflect amplification by the differential operator, but the 20.3% and 64% inter-output consistency defects still warrant comparison with this approach.

  3. Why is HTNO absent from Tables 12–13? How does it transfer across resolutions under the same budget, and how is the coarse vertex count or coarsening ratio set? For independent edge and face inputs, could circulation, flux, or higher-degree field reconstruction errors quantify information loss from vertex pooling and coarse-level reconstruction?

  4. On Darcy, which has scalar vertex outputs, how does the method compare with an established neural operator such as RIGNO? For Oriented-Darcy, a heterogeneous message-passing model retaining the same raw face data would provide a more direct DEC control than averaging those data first. Is such a comparison available?

  5. Why do the synthetic tasks use approximately 500–1,000 vertices and 100 meshes? Could one representative larger multi-degree mesh result show whether the component gains persist, together with preprocessing and inference costs?

5 comments

We thank the reviewer for recognizing the physical interpretation of our propagation paths and the strong public-benchmark results. We address the component controls, harmonic channel, and hierarchy below.

W1 / Q1 (a). Model configurations.

Which cochain degrees, spectral channels, copresheaf maps, and coarse levels are active in each public-benchmark model?

Across public benchmarks, TNO uses the same basic recipe. The model employs vertex–edge cochains, r=16r=16 retained spectral modes per degree (as discussed in gmQ3, W1 / Q1), fixed incidence transport with learned diagonal copresheaf maps, and no hierarchy. Elasticity additionally includes face cochains. HTNO uses vertex features with local updates mediated through edges and coarse correction, without spectral channels or learned incidence maps. The HTNO hyperparameters differ slightly across problem settings only in the number of V-cycle iterations (one to three) and of coarse regions (32–128), and further V-cycle iterations or subdivisions often did not improve the results. We have now included these additional details in Appendix F.

W1 / Q1 (b). Public-task component ablations.

What do public-task component ablations show with a common backbone, depth, and budget?

For the public benchmarks, all variants use the same 12-layer backbone with 128 width, and identical training protocol. We add either a spectral channel or a hierarchical correction through 32 geometry-based regions and four coarse message-passing layers for the hierarchical HTNO variant, which has no spectral channel.

Elasticity, mean ± SD over three seeds.

VariantSpectral channelCoarse correctionTest relative L1 (%)
LocalOffOff16.558 ± 0.187
Local + spectralOnOff2.774±0.110%\mathbf{2.774 \pm 0.110\%}
Local + hierarchyOffOn5.462±0.271‾\underline{5.462 \pm 0.271}

On Elasticity, hierarchy only partially compensates for the absence of the spectral channel, reducing the error from 16.558% to 5.462%, but remains 2.688 percentage points above the spectral variant (2.774%). Note that this is not equivalent to our HTNO model, which uses a deeper coarse correction and is designed for larger-scale tasks, but rather illustrates the effect of adding a single hierarchy to compensate for lack of spectral projection.

W1 / Q1 (c). Spectral truncation.

Are the eight modes a null space or a low-frequency subspace? What boundary conditions and truncation rule apply?

We incorporate positive modes alongside the kernel, please see our response to Reviewer gmQ3, W1 / Q1. We have now corrected this to r=16r=16 low frequency modes, and distinguished between the idealized TNO in Eq 8 and 9 and our experimental studies in sections 4 and 5 of the revised manuscript.

W1 / Q2 (a). Analytic harmonic control.

Do curl–curl labels satisfy the exact harmonic solution, and does TNO beat an analytic harmonic channel plus a learned complementary remainder?

The harmonic component of the curl–curl edge labels agrees with the analytic harmonic solution computed from the inputs to numerical precision (maximum discrepancy 4.3 × 10⁻⁷ over all 5,000 labels). To test the proposed control, we train the same TNO on the complementary remainder only and add the analytic harmonic component to its edge prediction. Model, split and training protocol are those of gmQ3, W1 / Q1.

Curl–curl relative L1 (%), mean ± SD over three seeds.

Harmonic edge componentVertex (%)Edge (%)Face (%)PDE residual (ratio)
Learned (full TNO)1.649±0.333\mathbf{1.649 \pm 0.333}2.358±0.114\mathbf{2.358 \pm 0.114}1.860±0.054\mathbf{1.860 \pm 0.054}5.181±0.154‾\underline{5.181 \pm 0.154}
Analytic, with learned remainder1.987±0.074‾\underline{1.987 \pm 0.074}2.526±0.066‾\underline{2.526 \pm 0.066}2.037±0.137‾\underline{2.037 \pm 0.137}4.685±0.274\mathbf{4.685 \pm 0.274}

Adding the analytic harmonic component does not reduce the errors, which are 0.17–0.34 pp higher than for the fully learned TNO, with the vertex difference small. Nevertheless, they lower the PDE residual from 5.18 to 4.68. The fully learned TNO therefore reaches comparable field accuracy when the harmonic component is not supplied in closed form. It receives the edge source and the low-frequency Hodge basis as inputs, which together enable learning this component.

W2 / Q4 (b). Heterogeneous message-passing control.

How does it compare with heterogeneous message passing using the same unaveraged face data?

We refer the reviewer to our response to Reviewer gmQ3, W2/Q2(a), where we compare DEC-based routing to a generic cellular message-passing architecture, both qualitatively and empirically. We ran this control on Darcy and curl–curl (Table 3b), and did not repeat it on Oriented-Darcy. On curl–curl, the cellular baseline receives the same unaveraged vertex, edge and face inputs as TNO, and DEC routing still has lower error on all three fields.

W1 / Q2 (b). Derived outputs and potential-only comparison.

Can deriving both outputs from AA enforce consistency, and how do field errors and PDE residuals compare?

To test this, we train a potential-only TNO that decodes only AA and sets q=δ1Aq=\delta_1A and B=d1AB=d_1A, which makes both identities exact by construction. We compare it with the TNO that decodes all three fields, under identical settings and three seeds. Both models train for 60 epochs to fit the rebuttal period, so the values below are not comparable with Table 3. We supervise the potential-only model on all three fields and weight the derived vertex and face losses so that the three terms have equal gradient norms at initialization. Without these weights, the derived losses dominate training and the model does not learn AA (71.6% edge error).

curl–curl, 60 epochsVertex rel. L1 (%) ↓Edge rel. L1 (%) ↓Face rel. L1 (%) ↓PDE residual ↓Vertex / face consistency defect (ratio) ↓
TNO, three decoded fields2.35±0.16\mathbf{2.35 \pm 0.16}4.89 ± 0.113.17±0.12\mathbf{3.17 \pm 0.12}14.17 ± 0.350.65 / 2.18
TNO, potential-only31.48 ± 0.024.49±0.25\mathbf{4.49 \pm 0.25}114.40 ± 1.648.02±0.09\mathbf{8.02 \pm 0.09}4.2×10−6/1.4×10−5\mathbf{4.2 \times 10^{-6} / 1.4 \times 10^{-5}}

The potential-only model satisfies both identities to floating-point accuracy, and it lowers the edge error and the PDE residual. However, its vertex and face errors rise from 2.35% and 3.17% to 31.48% and 114.40%, although both fields are supervised. We attribute this to the discrete derivatives, which amplify the remaining error in AA. The fully trained TNO of Table 3 (right) shows the same effect without retraining: setting q=δ1Aq=\delta_1A and B=d1AB=d_1A raises its vertex/face relative L1 from 1.60±0.33% / 1.87±0.07% to 20.60±0.74% / 65.05±2.08%. The twisted (copresheaf) transport maps provide learnable interactions between cochain features, increasing the flexibility of cross-rank DEC routing. They do not constrain the separately decoded outputs to satisfy q=δ₁A. Nevertheless, our experiments show that this structure provides a useful inductive bias for the neural models. Moreover, the compatibility induced by DEC transport could inspire new hybrid TNO variants, for instance, by projecting onto these constraints during a forward pass.

W2 / Q4 (a) / Q5. RIGNO on Darcy and Larger multi-degree meshes.

How does the method compare with RIGNO on Darcy? Why only 100 small meshes? Do gains persist on larger multi-degree meshes, including preprocessing and inference costs?

We evaluate our models across numerous public datasets, ranging from poisson problems, hyperelastic deformation, compressible airfoil flow from subsonic to supersonic regimes, and transonic 3D wing-surface aerodynamics, on meshes ranging from approximately 1K to 65K nodes. Across these benchmarks we observe overall improvements over RIGNO, in addition to faster convergence (see Figure 14), as well as several other baselines.

Our synthetic multi-rank controls (curl-curl and Darcy) are designed to carefully isolate the effects of the different architectural components, and substantiate the value of splitting dimensioned operators across topological ranks, using DEC transport. We keep these controls small, so that we can train every variant over three seeds, and evaluate larger meshes on the public benchmarks above. As we already compare to RIGNO across the public benchmarks, benchmarking RIGNO and other baselines on these synthetic examples would, in our eyes, overload what is already a dense manuscript.

W2 / Q3 (a). HTNO resolution transfer.

Why is HTNO absent from transfer tests, how does it perform at identical budget, and how is coarsening set?

We initially studied resolution transfer only for our spectral TNO instantiations (App. G). We have now run the same study for HTNO, our scalable TNO instantiation which mitigates the need for a spectral decomposition through hierarchical pooling (Fig 13.). HTNO is similarly trained at n=1,000n{=}1{,}000 and evaluated without retraining on the same held-out domains from App. G. The following reports curl–curl transfer, relative L1 (%), mean ± SD over three seeds. Bold denotes the better result for each vertex, edge, or face.

CochainModeln=500n{=}500n=1,000n{=}1{,}000 (train)n=2,000n{=}2{,}000n=4,000n{=}4{,}000
VertexTNO10.38±2.12\mathbf{10.38 \pm 2.12}1.84±0.42\mathbf{1.84 \pm 0.42}12.24±3.60\mathbf{12.24 \pm 3.60}45.78 ± 3.29
VertexHTNO11.79 ± 1.725.31 ± 0.0612.24±1.94\mathbf{12.24 \pm 1.94}33.26±12.84\mathbf{33.26 \pm 12.84}
EdgeTNO6.64±0.60\mathbf{6.64 \pm 0.60}2.06±0.41\mathbf{2.06 \pm 0.41}6.13±1.64\mathbf{6.13 \pm 1.64}15.59±1.07\mathbf{15.59 \pm 1.07}
EdgeHTNO20.91 ± 2.188.26 ± 0.1428.04 ± 0.1667.37 ± 1.97
FaceTNO5.22±1.02\mathbf{5.22 \pm 1.02}2.47±0.05\mathbf{2.47 \pm 0.05}7.40±0.27\mathbf{7.40 \pm 0.27}16.36±1.40\mathbf{16.36 \pm 1.40}
FaceHTNO9.10 ± 0.424.94 ± 0.3112.44 ± 0.5538.27 ± 0.75

On small meshes, the mass-weighted TNO variant (App. G) is the more principled model, as its spectral projection provides a powerful inductive bias (see gmQ3, W1 / Q1). This variant also uses no pooling and, as expected, has lower edge and face errors at every resolution. On vertices, HTNO transfers comparably, consistent with TNO at n=2,000n{=}2{,}000 and has lower error at n=4,000n{=}4{,}000 (33.26% against 45.78%). Its hierarchy coarsens only the vertices (see W2 / Q3 (b)), so edge and face features receive no coarse correction, which could be one reason for their degradation in resolution transfer. This design choice was conducted primarily because existing geometric PDE benchmarks do not require edge and face predictions. The hierarchy uses 32 coarse regions at every resolution. This number was fixed before the transfer evaluation and not tuned on the target meshes. We have now added these results to App. G.

We emphasize that we built the curl–curl meshes to study ablations in a controlled setting, whereas HTNO targets large meshes, for which spectral decomposition is costly (see App. G). Further improving HTNOs for discretization transfer would also be an important future enhancement. Nevertheless, both models still degrade under refinement, and we do not expect this to be easily or completely resolved. Zero-shot generalization across resolutions remains difficult even for neural operators designed for continuous functions (Bartolucci et al., 2023; Sakarvadia et al., 2026).

W2 / Q3 (b). Coarse transfers and pooling.

Does HTNO guarantee compatible transfer of higher-degree cochains? How much circulation, flux, and independent edge/face information is lost through pooling and reconstruction?

HTNO is a hierarchical extension of TNOs that uses pooling only as a coarse correction mechanism. Fine vertex features are preserved through a residual connection, and fine edge and face features are retained directly rather than reconstructed from pooled vertices, so pooling does not discard edge or face states.

The coarse representation does not preserve all within-region variation when constructed in this manner. Because the hierarchy is defined on combinatorial complexes, its transfer maps do not satisfy the de Rham commutation condition (as stated in App. D.1). HTNO therefore guarantees neither compatible transfer of higher-degree cochains between levels nor preservation of circulation or flux, but is a scalable alternative. We have now further clarified this in the appendix.

We hope that, given the additional evaluations and exposition, the reviewer considers raising their final score.

Reviewer_gmQ3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TNO represents fields as cochains on cell complexes and uses discrete exterior calculus (DEC) to couple different ranks. A hierarchical variant adds coarse-grid interactions. Validation covers irregular-mesh PDE benchmarks and synthetic Darcy/curl–curl problems, including ablations, resolution transfer, physical residuals, and computational costs.

Critical strengths
  1. The representation explicitly accommodates variables on vertices, edges, and faces, providing a physically motivated approach to coupled PDE fields.

  2. Broad experiments and parameter-matched controls support practical value. Reporting residuals, transfer failures, and overhead makes the evidence more informative.

Critical weaknesses
  1. The main performance-driving harmonic component is insufficiently specified. Equation (9) defines a full projection onto ker(Δ), whereas Appendix F.1 reports an r=8 harmonic-mode baseline. The exact retained spectrum and projection must be clarified to distinguish topological effects from generic low-frequency filtering.

  2. Key ablations confound DEC with other changes. Oriented-Darcy compares native face inputs with face-averaged graph inputs; Table 16 also changes copresheaf parameterization while removing cross-rank paths. These results do not isolate the claimed DEC-specific contribution.

Questions for authors
  1. Clarification: Which zero and nonzero modes are retained in each task, with what boundary conditions, inner product, and projection rule?

  2. Assessment-changing evidence: Does the advantage persist when baselines receive unaveraged face information and cross-rank ablations use identical copresheaf maps?

3 comments

We thank the reviewer for recognizing the physical motivation for posing neural operators in the language of DEC, the breadth of experiments, and the value of reporting residuals, transfer failures, and computational overhead. We address the spectral definition and the requested empirical studies below.

W1 / Q1. Harmonic and spectral modes.

Which zero/nonzero modes are retained per task, with which boundary conditions, inner product, and projection?

In Eqs. 8 and 9, we present an idealized TNO layer as derived through the Hodge decomposition. In our experiments, we observed that including a small number of positive modes improves performance, as we now present in the additional results below on our Darcy and curl-curl datasets. The full-spectrum TNO models thus use 16 Hodge-Laplacian eigenmodes per rank, including both zero and positive eigenvalues. We construct this basis from the input complex, independently of any specific PDE and its boundary conditions. At inference time, the learned operator remains agnostic to the underlying physics. The computed basis VV is orthonormal with respect to the Hodge stars MkM_k, which define the cochain inner products (Sec. 2), and the projection in Eq. 9 is the corresponding MkM_k-orthogonal projection. We study the experimental Darcy and curl settings from Table 3 (right):

Relative L1 (%) for the Darcy pressure and the curl–curl vertex, edge, and face fields, mean ± SD over three seeds.

Spectral channelDarcy (%)Vertex (%)Edge (%)Face (%)PDE residual (ratio)
Off7.472±0.0707.472 \pm 0.07017.569±0.35217.569 \pm 0.35216.004±0.02616.004 \pm 0.0266.757±0.0946.757 \pm 0.09424.427±0.65824.427 \pm 0.658
Nullspace only4.408±0.0274.408 \pm 0.0273.362±0.2713.362 \pm 0.2714.584±0.1184.584 \pm 0.1183.542±0.1183.542 \pm 0.1189.513±0.1749.513 \pm 0.174
Positive only2.234±0.009‾\underline{2.234 \pm 0.009}1.702±0.084‾\underline{1.702 \pm 0.084}3.396±0.116‾\underline{3.396 \pm 0.116}1.854±0.106\mathbf{1.854 \pm 0.106}6.845±0.558‾\underline{6.845 \pm 0.558}
Full retained spectrum1.888±0.017\mathbf{1.888 \pm 0.017}1.649±0.333\mathbf{1.649 \pm 0.333}2.358±0.114\mathbf{2.358 \pm 0.114}1.860±0.054‾\underline{1.860 \pm 0.054}5.181±0.154\mathbf{5.181 \pm 0.154}

The positive-only variant outperforms nullspace-only on all reported metrics, while the full setting yields the lowest Darcy and vertex/edge errors, as well as the lowest PDE residual. This supports that a spectral contribution beyond the harmonic subspace is useful. On curl–curl, we retrained all four variants for this study, and the full-spectrum row uses the same experimental setting as the full TNO of Table 3 (right). We now clarify in the main manuscript (sec. 4/5) that our TNO instantiations use more than just the kernel, despite the idealized presentation in Eq. 9. We maintain this formulation of an idealized TNO as part of the general framework we present, due to the Hodge derivation, which has now been additionally clarified with an example (lines 176-) In Appendix F.1 we errantly included r=8r=8 harmonic modes from a prior setup, which we have now corrected to the r=16r=16 used throughout.

W2 / Q2 (a). Comparison to Higher-Order Message Passing.

Does the advantage persist when baselines receive unaveraged face information?

Thanks to the reviewers' suggestions, we now additionally compare our proposed DEC routing with a generic learned cellular message-passing baseline (receiving the same unaveraged vertex, edge and face inputs) using identical encoders and decoders, and with the same global spectral channel. The baseline learns nonlinear messages from the source and target fields as well as the incidence sign. It follows the general message-passing scheme of simplicial and cellular networks (Bodnar et al., 2021a,b), restricted to the boundary and coboundary incidences that DEC routing uses, and Appendix F.7 now states its update rule (Eq. 23). In contrast, a TNO with DEC-routing is constructed through signed incidence sums and learns the subsequent channel transformations. The DEC rows use fixed incidence transport without learned copresheaf maps, i.e., the TNO of W1 / Q1 without these maps. Depth and training protocol are identical, and the models are parameter-matched on each dataset. We again study the experimental Darcy and curl settings from Table 3 (right). We ran this comparison on these two datasets and did not repeat it on Oriented-Darcy (Table 3a), so the evidence for unaveraged face inputs comes from curl–curl.

Relative L1 (%) for the Darcy pressure and the curl–curl vertex, edge, and face fields, mean ± SD over three seeds.

Spectral channelRoutingDarcy (%)Vertex (%)Edge (%)Face (%)PDE residual (median ratio)
OffGeneric cellular24.68±0.87‾\underline{24.68 \pm 0.87}34.52±1.41‾\underline{34.52 \pm 1.41}35.01±0.86‾\underline{35.01 \pm 0.86}18.78±1.49‾\underline{18.78 \pm 1.49}65.19±2.23‾\underline{65.19 \pm 2.23}
OffDEC7.86±0.27\mathbf{7.86 \pm 0.27}19.18±0.37\mathbf{19.18 \pm 0.37}18.47±0.39\mathbf{18.47 \pm 0.39}9.98±0.16\mathbf{9.98 \pm 0.16}31.11±3.97\mathbf{31.11 \pm 3.97}
OnGeneric cellular4.45±0.37‾\underline{4.45 \pm 0.37}3.86±0.13‾\underline{3.86 \pm 0.13}5.15±0.29‾\underline{5.15 \pm 0.29}4.59±0.28‾\underline{4.59 \pm 0.28}11.64±0.38‾\underline{11.64 \pm 0.38}
OnDEC2.29±0.02\mathbf{2.29 \pm 0.02}2.23±0.36\mathbf{2.23 \pm 0.36}2.87±0.08\mathbf{2.87 \pm 0.08}2.27±0.14\mathbf{2.27 \pm 0.14}6.17±0.33\mathbf{6.17 \pm 0.33}

With the spectral channel enabled, DEC reduces vertex/edge/face relative L1 errors from 3.86/5.15/4.59% to 2.23/2.87/2.27% (three-seed means of median errors over 500 test cases), and the Darcy error from 4.45% to 2.29%. Adding copresheaf maps with the same parameter budget yields the full TNO of W1/Q1. The accuracy advantage therefore persists when the baseline retains native higher-rank information and global communication, supporting our contribution of fixed DEC transformations in this comparison. These results are now included in Table 3b, with the full model definition in Appendix F.7.

Without the spectral channel, Cellular MP also has a higher error than the vertex-only MPNN in Table 3b (24.68% versus 8.59% on Darcy), and its Darcy validation error remains between 24% and 27% from epoch 50 onward across all three seeds. DEC routing computes the discrete gradient and divergence via fixed signed sums, whereas Cellular MP receives the sign only as input to a nonlinear message (Eq. 23) and must learn these operations from data.

W2 / Q2 (b). Controlled cross-rank ablation.

Does it persist when cross-rank ablations keep identical copresheaf maps?

We repeated the cross-rank ablation on curl–curl (the dataset and split of W1 / Q1 and W2 / Q2 (a)) with identical diagonal copresheaf maps and 10.413M parameters in both models, so that they differ only in whether cross-rank transport is enabled. Disabling it increases edge relative L1 from 2.35±0.12% to 32.42±2.10% (+30.07 pp), and vertex/face relative L1 from 1.46/1.80% to 99.18/45.40% (three-seed means ± sample SD of medians over 500 held-out cases). We have updated the cross-rank ablation in App. G to this comparison, which supports that cross-rank transport is a critical inductive bias for the proposed TNO instantiations.

We hope this clarifies the reviewer's questions and that, given the additional evaluations and exposition, the reviewer considers raising their final score.

Dear Reviewer gmQ3,

as the discussion period ends on November 18 (AoE), we would be grateful to hear whether the additional experiments and clarifications, namely the zero/positive-mode study (W1 / Q1), the cellular MP baseline with unaveraged face inputs (W2 / Q2 (a)), and the cross-rank ablation with identical copresheaf maps (W2 / Q2 (b)), have resolved your concerns. We are happy to clarify any remaining points in the meantime.

Best, The authors

Comments on the paper

1 comment

We thank the reviewers and the area chair for their time and valuable feedback. Reviewers gmQ3, 9Qbx, and hvUN highlighted the strong physical motivation for Topological Neural Operators (TNOs), which use discrete exterior calculus (DEC) to couple fields on vertices, edges, and faces. Reviewers 9Qbx and hvUN noted the compelling performance on public benchmarks, and 4GmY found the approach principled and effective across several regimes. Three of four reviewers reported learning something interesting or useful, and all four described their soundness concerns as minor or fixable.

The reviewers also asked for controls that isolate the roles of DEC routing and the spectral channel, time-dependent experiments, and a clearer account of how TNOs relate to prior work. We have addressed each reviewer comment with comprehensive, point-by-point responses and have summarized the most significant additional results and revisions below.

Generic MP, DEC Routing and Spectral Channels. Comparing to a generic cellular message passing scheme with identical inputs and spectral channels at comparable parameter counts, DEC routing reduces steady curl–curl edge relative L1 error from 5.15% to 2.87% and face error from 4.59% to 2.27% over three seeds (now in Tab. 3b, gmQ3 W2/Q2(a)). Without cross-rank transport, edge error rises from 2.35% to 32.42% (Table 16, gmQ3 W2/Q2(b)). The spectral channel contributes substantially, and the new controls show that positive Hodge modes account for a significant part of its improvement (gmQ3 W1/Q1). We clarify these points in the main manuscript, including the distinction between exact DEC identities and conservation of learned predictions (hvUN W6(b)). The updated manuscript now also reports resolution transfer for the hierarchical model (HTNO) (9Qbx W2/Q3(a)), forward pass costs for key components (hvUN W1(c)/Q2), the Elasticity component ablation with error bars (9Qbx W1/Q1(b)), the analytic harmonic control (9Qbx W1/Q2(a)), and the potential-only study (9Qbx W1/Q2(b)), as requested in 9Qbx’s follow-up.

Clarifying the exposition. We added a main-text related-work section (4GmY W1) and an example in the preliminaries on a small mesh (hvUN W7). We added a guide at the start of the appendix to help readers locate the derivations and experiments (hvUN W7). We also clarified the comparison with HSD, which we compare extensively against in Table 6, and TNO's relationship to Geo-NeWF (4GmY W2).

Experimental Scope. Learned steady-state solution operators are becoming a central component of many engineering systems across academia and industry. Our extensive experiments cover established public benchmarks in elliptic PDEs, elasticity, and compressible flow (Tables 1–2). Reviewer hvUN appears to insist this work also addresses the time-dependent setting. We would like to stress that learning steady-state solutions of PDEs is already a challenging task and a central component of many engineering systems across academia and industry, and has been the sole focus of state-of-the-art works such as GINO, Geo-NeWF, AB-UPT, and large-scale benchmarks such as Emmiwing. Our focus in this work is not dynamics but the domain. We bring insights from discrete exterior calculus to the world of neural operators and propose a novel, general class of architectures. We kindly ask that our work be assessed on its merits within its scope. Time-dependent dynamics on meshes warrant treatment in their own right, and we appropriately leave this for a future study. We emphasize that this intentional restriction of scope does not limit the generality of our framework, which applies to any map between cochain spaces, and should not, on its own, be grounds for rejection.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.