acceptodds
Under review as a conference paper at ICLR 2027

Beyond Arbitrary Geometry: Topology Generalization In neural PDE Operators

20 views

Abstract

Neural operators that accept arbitrary meshes are often treated as geometry-general, but unseen domain topology changes both the invariant and decaying subspaces of a PDE operator. We use Hodge heat flow as a controlled lens on this distinction and introduce TopoBox-3D, where tunnels and cavities vary Betti support while the exact Hodge decomposition separates the harmonic kernel from the positive spectrum. Across six architectures, models that infer topology implicitly suffer excess matched degradation in 37 of 45 model–task topology-OOD cells, yet cases that change harmonic dimension are not more strongly penalized on average. The dominant difficulty is instead spectral: the initial Rayleigh quotient is the most stable predictor of error, and spectral broadening adds information for edge and face cochains. Most strikingly, controlled probes show that explicit incidence and harmonic coordinates do not yield the best kernel-identity accuracy; nevertheless, TNO ranks first in mixed-input nonharmonic accuracy on all six tasks with nontrivial harmonic support. Together, these results establish topology as a distinct generalization axis beyond arbitrary-geometry compatibility and show that its influence extends across the Hodge spectrum rather than remaining confined to the harmonic kernel. More broadly, they suggest that global, low-frequency structural priors may help organize predictions in the faster-decaying complementary component, offering a new perspective on how neural operators may generalize across topology as well as geometry.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Neural Operators, Topology Generalization, Scientific Machine Learning, PDE Surrogate Modeling, Hodge Decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VzWiOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates generalization to unseen domain topology in neural PDE operators. It introduces TopoBox-3D, a benchmark with independently varying tunnels and cavities, and evaluates six architectures on Hodge heat diffusion. The analysis compares geometry and topology shifts, examines associations between spectral descriptors and prediction error, and measures harmonic preservation and nonharmonic-component accuracy.

Critical strengths

The motivation is relevant: compatibility with arbitrary meshes does not establish generalization to topological configurations absent from training. The independently controlled tunnels and cavities provide a concrete experimental setting, and Figure 1(a) helps communicate the domain families. The reported topology audits, geometry-level data splits, and comparisons across several architectures are useful features of the benchmark.

Critical weaknesses

My main concern is that the benchmark covers a narrow PDE setting and the experimental analysis provides limited explanation of why models perform poorly under particular geometric or topological changes. The following issues limit the significance of the contribution.

  1. The implications beyond Hodge diffusion remain unclear. All three cochain degrees belong to the same linear, autonomous diffusion family, evaluated at a fixed time under homogeneous absolute boundary conditions. Hodge diffusion is a reasonable diagnostic choice because its harmonic kernel has a direct relationship to topology and its spectral evolution is analytically tractable. However, this special structure also limits what the experiments establish about neural PDE operators more broadly. The scalar, circulation, and flux interpretations do not demonstrate applicability to distinct physical PDE systems. Section 6 acknowledges these limitations, but I would like a clearer account of which findings are expected to transfer to other operators or boundary conditions, and why. At present, the broader implications are primarily hypotheses.

  2. The heterogeneous results need a stronger explanatory interpretation. Section 5.2 reports that 37 of 45 comparisons for the implicit-topology models have (Q>1), but several cells show (Q<1). These indicate less degradation than the fixed-topology geometry baseline. Separately, Table 8 contains cases with (R<1), indicating lower aggregate OOD error than IID error. These exceptions do not invalidate the study, but they raise an important question: under what conditions does an unseen topology make prediction harder, leave it largely unaffected, or coincide with an easier test distribution? The aggregate tendency does not answer this question. Since topology, boundary structure, mesh size, and spectrum change together, the results need more analysis connecting particular domain changes to the observed model behavior.

  3. The diagnostics describe associations more convincingly than they explain model failures. The Rayleigh-quotient correlations and component-level errors provide useful summaries, but they do not establish why particular architectures succeed or fail in different settings. The paper itself notes that the association with relative error may partly reflect target attenuation in the error denominator. Likewise, TNO’s nonharmonic advantage cannot be attributed separately to incidence structure or harmonic information. These qualifications are appropriately acknowledged, but they leave the proposed architectural explanation unresolved. My concern is therefore whether the analysis yields sufficiently actionable insight beyond documenting performance differences within this benchmark.

  4. The mathematical exposition makes the contribution difficult to assess. The introduction assumes considerable familiarity with Betti numbers, homology/cohomology, differential forms, and Hodge decomposition. Section 3 then introduces tetrahedralization, chain and cochain complexes, boundary matrices, matrix ranks, and the Euler characteristic with little conceptual explanation. Although Appendices A and B provide additional details, the main text needs a clearer progression from the geometric construction to its topological invariants and then to the diffusion operator. In particular, explaining why removing through-cylinders and enclosed balls gives the stated Betti vector would help readers understand the benchmark. Equation (1)’s kernel–cohomology relationship also deserves a concrete example illustrating harmonic preservation and nonharmonic decay.

Questions for authors
  • Why hodge diffusion is chosen in the benchmark? Which conclusions are specific to Hodge heat diffusion, and what evidence or mathematical argument supports their relevance to other PDEs?

  • Can the authors explain representative (Q<1) and (R<1) cases using the existing geometric, spectral, and component-level diagnostics? What distinguishes them from settings with substantial degradation?

  • What specific mechanism explains the differences between architectures? Could an analysis using absolute or input-normalized error, or a focused structural ablation, strengthen that interpretation?

5 comments

Thank you for the detailed review. We added normalization and composition controls, geometric matching, and a three-seed subspace audit to address your questions, and clarified the mathematical exposition.

1. What explains the cases with Q<1Q<1 or R<1R<1?

The inequalities answer different questions: RpR_p is the median seed-wise OOD/IID mean-error ratio, whereas Qp=Rp/RAQ_p=R_p/R_A compares degradation against Protocol A.

For MGN D2, R=1.129R=1.129 but Q=0.815Q=0.815: error increases, but less than in A2 (RA=1.385R_A=1.385). Thus, Q<1Q<1 does not imply improved OOD accuracy.

For TNO C2, normalization changes the conclusion:

\text{Error normalization} & R & 95\\%\text{ geometry interval} \\\\ \hline \text{Target norm} & 0.776 & [0.737,0.822] \\\\ \text{Input norm} & 1.035 & [1.019,1.052] \end{array}$$ Target-relative $R$ is 1.059 on nonharmonic inputs, versus 0.738, 0.515, and 0.419 for weak, balanced, and strong harmonic mixtures. Standardizing realized harmonic-fraction composition on common support gives $R=1.080$; adding Rayleigh-quartile stratification gives 0.949. Geometric matching yields 1.016 and 1.015, with both intervals including one. These are descriptive contrasts on restricted populations. **They show that aggregate OOD improvement depends on composition and normalization**, rather than establishing intrinsically easier unseen topology (Sections 5.2–5.3; Appendices D.3–D.4). ### 2. Absolute/input-normalized error control Following your suggestion, we use $$E_{\mathrm{in}}=\frac{\\|\widehat w_T-w_T\\|_M}{\\|w_0\\|_M}, \qquad q_T=\frac{\\|w_T\\|_M}{\\|w_0\\|_M}.$$ Ignoring the small numerical stabilizer, $E_{\mathrm{rel}}=E_{\mathrm{in}}/q_T$. Inputs have unit mass norm to numerical precision, so absolute and input-normalized errors coincide. The median Rayleigh–error correlations are: $$\begin{array}{lrrrr} \text{Normalization} & k=0 & k=1 & k=2 & \text{All 72 cells} \\\\ \hline \text{Target} & .837 & .729 & .783 & .784 \\\\ \text{Input} & .770 & -.254 & -.433 & .024 \end{array}$$ **This confirms the importance of target attenuation.** We revised the difficulty interpretation accordingly. The main finding remains: input normalization gives $Q>1$ in **41/45** implicit-model cells, with geometric mean 1.183. The control refines the explanation while preserving the widespread generalization gap (Section 5.4; Appendix D.2). ### 3. What can be concluded about architectural mechanisms? The expanded audit uses three frozen checkpoints and ten geometries per split across six tasks and six models: **2,160 evaluations**. It identified a sparse-input construction fault in the earlier TNO probes. We corrected and reference-verified the inputs and recomputed all 360 TNO records. Corrected OOD results are: $$\begin{array}{lrr} \text{Model} & I_H & L_{H\to\perp} \\\\ \hline \text{MGN} & .1887 & .1279 \\\\ \text{RIGNO} & .1874 & .1088 \\\\ \text{Transolver} & .2910 & .1473 \\\\ \text{GNOT} & .1984 & .1386 \\\\ \text{GAOT} & .5913 & .0538 \\\\ \text{TNO} & \mathbf{.0231} & \mathbf{.0226} \end{array}$$ Here $I_H$ is total harmonic-input error and $L_{H\to\perp}$ outward leakage. These are pooled medians of three-seed medians over 60 unit-mass task–geometry probes per model; some tasks share geometries. **We withdrew the earlier harmonic/complementary tradeoff claim:** TNO now shows low harmonic error and still ranks first in complementary accuracy on all six tasks. We verified 24 incidence matrices; a 192-sample conventional-inference spot check agrees with stored relative errors within $4.36\times10^{-5}$, without constituting a full benchmark rerun. Appendix D.5 provides task-wise intervals and full-test nonharmonic leakage. The evidence supports TNO's combined structural access; separating incidence from harmonic information requires a focused ablation. Zero/identity baselines show why harmonic preservation alone is insufficient to establish topology recognition (Appendix D.6). ### 4. Hodge diffusion and clearer exposition Hodge heat links topology to testable operator identities. The Introduction now illustrates preservation versus decay: $$w_0=a h+b u_\lambda \quad\Longrightarrow\quad w_T=a h+b e^{-\kappa T\lambda}u_\lambda,$$ Here $h$ is harmonic and $u_\lambda$ has positive eigenvalue. Before introducing the chain complex, Section 3 explains $(1,g,c,0)$ through connected material, tunnel loops, cavity cycles, and the absence of a closed three-dimensional cycle. Section 6 separates reusable evaluation lessons from Hodge-specific identities and rankings; the three degrees remain one PDE family. We hope these controls and clearer explanations strengthen the empirical contribution and address your concerns.

Thank you for the detailed response and substantial revision. Several of my earlier concerns have been addressed. Sections 5.2–5.3 now clarify the distinction between Q<1Q<1 and R<1R<1, and the composition and matching analyses help explain why some aggregate OOD results appear favorable. The input-normalized error controls address the missing normalization check and appropriately limit the interpretation of the Rayleigh–error association. I also appreciate the correction of the TNO probe construction, the expanded audit across three checkpoints, and the withdrawal of the earlier harmonic/complementary tradeoff claim.

However, the mathematical background remains insufficiently developed for readers without prior training in topology and Hodge theory. The added examples provide useful intuition, but essential concepts—including Betti numbers, homology/cohomology, differential forms, chain/cochain complexes, and harmonic subspaces—still appear without a coherent introduction. Section 2 primarily discusses related work. A concise background subsection explaining these concepts and their relationships before the central equations would substantially improve accessibility. A worked example connecting a tunnel or cavity to a nonbounding cycle, its Betti number, and the corresponding harmonic space would be particularly helpful. These foundations are necessary for understanding what the benchmark measures and why its diagnostics are meaningful.

My other main concern is what actionable guidance the benchmark provides for improving models. The revisions strengthen evaluation practice through better treatment of normalization and input composition. However, the connection between the diagnostics and specific model improvements remains unclear.

For example, TNO’s corrected performance suggests that its combined structural access is promising. The comparison does not establish which ingredient contributes to that advantage: incidence information, harmonic coordinates, processor design, or their interaction. Consequently, it is difficult to infer which change a practitioner should make to an existing model after observing poor performance on this benchmark.

Could the authors clarify two points?

  1. How can the benchmark guide a concrete intervention? Can a particular diagnostic—such as harmonic distortion, cross-subspace leakage, or complementary-component error—identify a model modification likely to improve OOD performance? A focused ablation or mechanistic argument connecting one diagnosed failure to one intervention would strengthen the benchmark’s value for model development.

  2. Under what conditions might that intervention benefit other PDEs? Section 6 appropriately acknowledges that the numerical results and harmonic identities cannot be assumed to transfer. It would nevertheless be helpful to explain which proposed improvements depend on Hodge diffusion’s particular structure and which have a plausible basis for improving geometry/topology OOD generalization in other PDE settings.

I recognize that a diagnostic benchmark can be valuable without introducing a new architecture. My remaining concerns are whether the presentation gives readers the background needed to understand the diagnostic framework, and whether the evaluation supplies sufficient guidance for model improvement with potential relevance beyond this operator family.

Thank you for the careful follow-up and for acknowledging the improvements. We have revised Sections 2.1, 5.6, and 6 to connect the mathematical background, a concrete model intervention, and its conditions for transfer.

1. A connected background and worked example

Section 2.1 now provides a connected primer and worked example before the first Hodge-operator equation; related work follows in Section 2.2. The sequence is:

Chains and cycles → homology and Betti numbers → cochains and differential forms → harmonic representatives.

A chain is a weighted collection of oriented simplices; its boundary is given by an incidence matrix. A cycle has zero boundary. Homology identifies cycles differing by a boundary, and the Betti number counts independent classes over the reals. Cochains assign linear measurements to chains; integration of differential forms gives such measurements. The transpose incidence maps form the cochain complex. Cohomology identifies closed cochains modulo exact ones; with the mass inner product and our boundary conditions, each class has a unique harmonic representative.

Worked example: a box with one through-tunnel and no cavities has one independent loop that cannot bound a surface within the material, so β1=1\beta_1=1. Its dual cohomology class has a harmonic 1-form representative hh with nonzero circulation around that loop. Thus Δ1h=0\Delta_1h=0 and heat evolution preserves hh. By contrast, removing an enclosed ball produces a nonbounding surrounding surface, contributing to β2\beta_2 and the harmonic 2-form space. This connects the geometric construction directly to the stationary components being audited.

2. A concrete intervention with a pointwise error guarantee

Section 5.6 adds harmonic-output correction with a pointwise guarantee (Equations 7–8). For any frozen predictor FF, define

F~(w0)=PHw0+(I−PH)F(w0).\widetilde F(w_0)=P_Hw_0+(I-P_H)F(w_0).

This replaces only the predicted harmonic component by its known, conserved value. The projector is computed from the test-domain discretization, without test solution labels. Since PHwT=PHw0P_Hw_T=P_Hw_0, mass orthogonality gives

=\\|F(w_0)-w_T\\|_M^2 -\\|P_H(F(w_0)-w_T)\\|_M^2.$$ **The diagnosed harmonic error is exactly the removable squared error.** With an exact projector and conserved target component, the correction cannot increase absolute or target-/input-normalized error on any IID or OOD input. Improvement is strict when the original harmonic error is nonzero; the degradation ratios $R$ and $Q$ need not decrease. This yields a specific decision rule: - **Substantial harmonic error or complement-to-harmonic leakage:** apply the correction; the harmonic error energy measures its recoverable benefit. - **Predominantly complementary error:** this correction offers little gain, directing attention to the learned complementary dynamics instead. It removes within-harmonic distortion and complement-to-harmonic leakage, while leaving complementary error, including harmonic-to-complement leakage, unchanged. It requires additional structural information for models not already given $P_H$; its additional basis and projection costs must also be reported. This is an explicit mechanistic guarantee, not a new empirical ablation or a claim to isolate the ingredients behind TNO's advantage. Section 5.6 explains both when this modification helps and when it cannot address the dominant error. ### 3. Conditions for transfer to other PDEs Section 6 states the transfer condition (Equation 9): **enforce a known solution component**. Suppose a domain-specific orthogonal projector $P$ satisfies $$Pw_T=g(w_0,f,\mathrm{BC}),$$ where $g$ is known from the initial data, forcing, and boundary conditions. Then $g+(I-P)F(w_0)$ obeys the same error-reduction identity. This does not require the full PDE to be linear: a known mass balance, for example, can determine the constant component in a conservative system. For Hodge heat, $P=P_H$ and $g=P_Hw_0$. For another PDE, the relevant component and its evolution must be established from that equation and its boundary conditions; blindly preserving a Hodge component is not justified. Approximate projectors or an incorrect balance law also remove the exact guarantee. This provides a concrete route from diagnosed invariant-component error to model correction, with explicit conditions for broader applicability. We hope this mechanism and the revised background address the remaining concerns about accessibility and model-development value.
Reviewer_TdDqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TopoBox-3D, a benchmark for evaluating how well neural PDE operators generalize to unseen domain topology. The benchmark constructs 3D meshes with varying tunnels and cavities and uses Hodge heat flow to study topology generalization on vertex, edge, and face cochains. The authors evaluate six neural operators under geometry-OOD and topology-OOD settings and find that most models degrade further under topology shifts. They also observe a strong correlation between prediction error and spectral properties, particularly the initial Rayleigh quotient, and analyze how different architectures preserve harmonic components and predict nonharmonic components.

Critical strengths
  • The paper studies an interesting problem by distinguishing topology generalization from geometry generalization in neural PDE operators.
  • The benchmark is carefully constructed, with controlled topology variations and clearly defined evaluation splits.
  • The experiments cover a range of neural operator architectures, with useful spectral and component-wise analyses beyond overall prediction accuracy.
Critical weaknesses
  • While I agree that topology generalization is worth studying, I am not fully convinced of the need for a new benchmark. The paper focuses more on analyzing model performance than establishing what the proposed benchmark offers beyond existing evaluations. For a benchmark paper, I would expect a clearer discussion of its distinctive features, practical utility, and advantages over prior benchmarks. Although the paper provides dataset statistics, a more systematic characterization across different topology classes would be helpful. I would also like to see some discussion of computational cost, considering that the benchmark involves mesh generation, Hodge decomposition, and PDE simulation.
  • Some observations, such as the strong correlation between the initial Rayleigh quotient and prediction error, are reasonably expected for a diffusion problem. Other findings, such as the lack of additional degradation when the harmonic dimension changes, are more interesting but are not explored in sufficient depth. Similarly, the architectural comparisons reveal different performance patterns, but it is unclear what general lessons can be drawn beyond the specific models and PDE considered. I wonder how these findings could inform future research on topology generalization.
  • The experiments are limited to a single linear Hodge heat equation with a fixed evolution time and boundary conditions. The topology variations come from a relatively simple family of synthetic domains.
Questions for authors
  • Could the authors report prediction errors separately for the four initial-condition energy mixtures? I am interested in whether the relative performance of different architectures changes with the harmonic energy fraction.
3 comments

Thank you for recognizing the problem's importance and the controlled benchmark design. We have added the four-mixture breakdown, topology-class statistics, and recorded computational costs to clarify the benchmark's practical value.

1. What does TopoBox-3D add as a benchmark?

Its distinctive value is the combination of controlled topological support and an auditable operator. Independent tunnel/cavity variation specifies which topological configurations are absent from training; geometry-level splits prevent leakage between initial conditions on the same domain; and the Hodge decomposition provides known identities for assessing harmonic preservation and complementary prediction.

This enables tests of unseen-topology transfer, mixture-dependent OOD improvements, and stationary versus decaying component accuracy. The Introduction and Section 3 clarify this role. Appendix D.7, Table 25 adds topology-class statistics for geometry count, volume, boundary area, vertex count, and mesh quality.

2. Does architecture ranking depend on the four energy mixtures?

We have added the requested complete breakdown: Tables 15–18 report all six models, all 12 tasks, both test splits, and all four requested mixtures. Each entry averages the 200 geometries within a seed and then takes the median across three seeds.

The requested harmonic fractions are 0,1/9,1/3,1/20,1/9,1/3,1/2. We explicitly distinguish these from realized fractions: when the relevant harmonic space is absent, unavailable harmonic energy is redistributed; scalar inputs are mean-zero.

For illustration, the OOD relative errors on C2 are:

\text{Model} & 0 & 1/9 & 1/3 & 1/2 \\\\ \hline \text{MGN} & .7203 & .4949 & .3507 & .3029 \\\\ \text{RIGNO} & .5338 & .4892 & .4064 & .3848 \\\\ \text{Transolver} & .7016 & .7760 & .6088 & .5127 \\\\ \text{GNOT} & .6174 & .6983 & .5440 & .4484 \\\\ \text{GAOT} & .8317 & .7870 & .7786 & .7739 \\\\ \text{TNO} & \mathbf{.3216} & \mathbf{.1292} & \mathbf{.0697} & \mathbf{.0509} \end{array}$$ This reveals both stability and variation. TNO has the lowest point estimate in **all 24 OOD task–mixture combinations** across the six tasks with nontrivial harmonic support. Rankings among the other models do change: on C2, RIGNO outperforms MGN for nonharmonic input, whereas MGN outperforms RIGNO for the balanced and strong harmonic mixtures. We report these as point-estimate rankings; paired intervals are used where discussing separation between close models. The breakdown also explains why aggregate degradation can be misleading. For TNO C2, the OOD/IID relative-error ratios are: $$\begin{array}{lrrrr} \text{Requested harmonic fraction} & 0 & 1/9 & 1/3 & 1/2 \\\\ \hline R & 1.059 & 0.738 & 0.515 & 0.419 \end{array}$$ Its aggregate $R<1$ consequently conceals a different behavior on purely nonharmonic inputs. Section 5.3 and Appendix D.3 now develop this example and include realized-composition controls. ### 3. What lessons go beyond expected diffusion behavior? We agree that a Rayleigh–error correlation alone would offer limited new insight. The new normalization control changes the interpretation: across 72 model–task cells, its median correlation falls from **0.784** with target-relative error to **0.024** with input-normalized error. Yet excess topology-OOD degradation remains widespread: $Q>1$ occurs in **41/45** implicit-model cells under input normalization, compared with 37/45 under target normalization. These results motivate a concrete evaluation practice: report both normalizations, audit realized harmonic support, and distinguish an OOD/IID ratio from its geometry-reference contrast. They also refine the architectural evidence. After correcting a sparse-input construction issue in the earlier TNO probes, the expanded three-seed audit finds low TNO harmonic-input error ($I_H=0.0231$), together with first-place complementary-component accuracy on all six supported tasks. Section 5.5 replaces the earlier capability-asymmetry interpretation; Appendix D.5 documents the correction and validation. ### 4. Computational cost and scope Appendix D.7, Table 26 now reports recorded training and OOD-inference times for every model and degree, summarized over 12 runs per cell. These are measured log timings; separate mesh-generation, Hodge-preprocessing, and PDE-generation times were not recoverable and are not included. We have also made the role of Hodge heat explicit: its tractable kernel and decay structure enable these controls, while the three cochain degrees remain one PDE family. Section 6 separates the reusable evaluation lessons from conclusions that require additional PDE validation. We hope these additions clarify the benchmark's utility and its lessons for topology-generalization research.

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' effort in addressing my concerns and improving the manuscript.

My concerns regarding Q1 are largely addressed, and I appreciate the clarification on W1 and W3. However, the revised W2 results weaken some of the original conclusions, particularly the Rayleigh-error correlation and the architectural capability asymmetry. I still find it difficult to identify insights general enough to extend beyond the specific models and linear Hodge heat problem considered.

Overall, I think the authors have made a substantial effort in the rebuttal, and the paper has improved. But I will maintain my rating at 2 for now. I will also consider the feedback from other reviewers during the remaining discussion period and adjust my rating if warranted. As a side request, could the authors mark the additions in the revised manuscript in a different color, i.e., blue? That would help me assess the improvement of the manuscript more closely.

Thank you for your continued assessment and for recognizing the improvements. We appreciate that your question about the energy mixtures is largely resolved. Following your suggestion, we have prepared a revised manuscript with added or revised text marked in blue, relative to the originally reviewed version. Unchanged text remains black, including material that was only relocated; the original plotting colors are preserved.

We agree that the normalization control and corrected TNO audit require narrowing two earlier interpretations. The revised paper no longer relies on a universal Rayleigh–difficulty association or on the withdrawn harmonic/complementary capability tradeoff. Section 6 now distinguishes three conclusions that remain supported:

  1. Empirical robustness within the benchmark. Excess topology-OOD degradation persists after removing target normalization: Q>1Q>1 in 41/45 implicit-model cells, with geometric mean Q=1.183Q=1.183 (Section 5.2; Table 13). The normalization control changes the proposed explanation but preserves this central empirical finding.

  2. A reusable evaluation principle. The identity Erel=Ein/qTE_{\mathrm{rel}}=E_{\mathrm{in}}/q_T (ignoring the numerical stabilizer) applies beyond Hodge heat. It explains why target-relative error can conflate approximation error with attenuation. Sections 5.3–5.4 show how normalization and input composition can change the interpretation of apparent OOD improvements, motivating controls applicable to other operator-learning evaluations.

  3. A concrete, conditional model intervention. Section 5.6 gives harmonic-output correction and its pointwise error identity (Equations 7–8). Section 6 states the broader condition: when a known solution component satisfies PwT=g(w0,f,BC)Pw_T=g(w_0,f,\mathrm{BC}), replacing that component by gg gives the same nonincrease guarantee. This is a mechanistic argument under explicit assumptions, rather than empirical validation on additional PDE families or an attribution of TNO's advantage to one ingredient.

For convenient review, the blue-marked changes also include the connected mathematical primer in Section 2.1, the four-mixture results in Tables 15–18, the corrected probe audit in Section 5.5 and Appendix D.5, and the distribution and cost summaries in Appendix D.7.

We understand your remaining concern about broader significance. Our intended contribution is a controlled benchmark that exposes these evaluation and model-diagnostic issues, with a clear separation between findings demonstrated within Hodge diffusion and principles that apply under more general conditions. We hope the marked manuscript and this distinction make the strengthened contribution easier to assess. Thank you again for your time and constructive feedback.

Reviewer_Gx68OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether neural operators that can handle arbitrary geometries can also generalize to domains with unseen topologies. The paper introduce a synthetic benchmark with different numbers of tunnels and cavities. They use the heat equation to study how topology changes affect the performance of neural operators. In the experiments, they compare different architectures under different geometry and topology shifts. They show that unseen topology can introduce additional errors compared with geometry changes alone. They also investigate how harmonic and nonharmonic components contribute to prediction errors.

Critical strengths
  • The paper studies an interesting problem that is not sufficiently addressed in many neural operator benchmarks. The distinction between handling arbitrary geometries and generalizing to unseen topologies is important. The authors also provide a relatively large synthetic benchmark. Comparing different architectures across several generalization settings makes the experimental study informative.
  • The analysis goes beyond reporting only prediction error. In particular, separating harmonic and nonharmonic components and investigating their relationships with prediction errors provides useful insights. The authors also use multiple training seeds and report uncertainty for their comparisons.
Critical weaknesses
  • One of my main concerns is the limited generality of the experiments. All experiments are performed on the same linear heat equation with fixed boundary conditions and a fixed prediction time. Although the authors consider domains with different numbers of tunnels and cavities, the geometries are still limited to a synthetic box based setting. Also, in the OOD experiments, the authors mainly change the number of tunnels or cavities, while more substantial geometric changes are evaluated separately. It is unclear whether the conclusions would hold for more realistic PDEs, such as nonlinear problems or problems with more complicated boundary conditions. Although the paper acknowledge these limitations, I believe an experiment with another PDE would significantly strengthen the contribution of the paper.

  • Another concern is that the experiments do not fully separate the effect of topology from other factors. For example, changing the number of tunnels or cavities also changes the boundary geometry, mesh size, and spectrum of the operator, so it is unclear how much of the additional error is directly related to topology.

Questions for authors
  • How would the conclusions change for nonlinear PDEs or more realistic domain geometries? In particular, would the additional topology-OOD degradation remain similar when the underlying dynamics are different from Hodge heat diffusion?
  • Can the authors better isolate the effect of topology from changes in mesh size, boundary geometry, and spectral properties? For example, would it be possible to construct more closely matched domains that differ in topology but have similar discretization statistics?
1 comment

Thank you for recognizing the importance of distinguishing arbitrary-geometry compatibility from generalization to unseen topology, and for the constructive suggestions on generality and confounding. We have added geometric matching, normalization controls, and initial-condition analyses using the existing test data and frozen checkpoints. These additions strengthen the empirical finding while making its interpretation more precise.

1. Why Hodge diffusion, and what can transfer beyond it?

Hodge diffusion provides a useful diagnostic because it makes the relationship between domain topology and operator behavior explicit:

S_T=P_H+e^{-\kappa T\Delta_{k,\perp}}P_\perp.$$ This gives two independently measurable requirements: preserving the harmonic component and predicting decay on its complement. A change in prediction error can therefore be examined against known operator identities, beyond an aggregate accuracy comparison. **The transferable contribution is the evaluation framework:** explicitly withholding topological support, separating geometry-level splits, and checking whether conclusions survive changes in initial-condition composition and error normalization. The magnitude of degradation on a nonlinear PDE cannot be inferred from the present experiment. We now distinguish this general evaluation principle from the Hodge-specific identities and empirical rankings, and explicitly describe the three cochain degrees as one operator family. The revised Introduction and Sections 3 and 6 make this scope clear. ### 2. Can topology be better separated from mesh and boundary geometry? Following your suggestion, we added **optimal one-to-one geometric matching** within each protocol and degree. Matching uses log token count, log volume, log boundary area, and minimum mesh quality, with a caliper imposed on every covariate. The numbers of retained pairs are: $$\begin{array}{lcc} \text{Protocol} & 0.5\text{-SD caliper} & 1.0\text{-SD caliper} \\\\ \hline A & 28\text{--}30 & 95\text{--}98 \\\\ B & 0 & 12 \\\\ C & 25 & 64\text{--}65 \\\\ D & 0 & 11\text{--}12 \end{array}$$ Each split contains 200 geometries; ranges are across cochain degrees. We report matched error contrasts only when at least 20 pairs remain, and report residual covariate imbalance alongside coverage. This control is informative for the apparent TNO C2 improvement. Its unmatched target-relative OOD/IID ratio is $R=0.776$. After matching, it becomes $1.016$ with 95% interval $[0.933,1.110]$ at the stricter caliper, and $1.015$ with interval $[0.967,1.064]$ at the wider caliper. Thus, the aggregate improvement does not persist as a clear improvement in these geometrically comparable subsets. The limited overlap in B/D is also an explicit result of the analysis: we do not extrapolate a matched effect where comparable geometries are absent. **The revised paper describes topology-associated distribution shifts and uses Protocol A as a reference contrast, without attributing the entire excess error to Betti numbers alone.** Details are in Section 5.2 and Appendix D.4, Table 19. ### 3. Does the main finding survive an error-normalization control? Yes. We recomputed degradation using input-normalized error, $$E_{\mathrm{in}}=\frac{\\|\widehat w_T-w_T\\|_M}{\\|w_0\\|_M}, \qquad Q=\frac{R_p}{R_A}.$$ Across the five implicit-topology models and nine topology-OOD tasks: $$\begin{array}{lcc} \text{Error normalization} & \text{Cells with }Q>1 & \text{Geometric mean }Q \\\\ \hline \text{Target norm} & 37/45 & 1.110 \\\\ \text{Input norm} & 41/45 & 1.183 \end{array}$$ **The widespread excess degradation therefore survives removal of the target-norm denominator.** At the same time, the matching analysis identifies where a more specific attribution is supported and where it remains unresolved. All intervals resample geometries while retaining their initial conditions and the three fixed checkpoints. We have incorporated these results in Section 5.2 and Appendix D.1, Table 13. We hope these additional controls address your concern about separating the observed generalization gap from the factors that accompany topology change, and make the benchmark's contribution more convincing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.