acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Kuramoto–Hodge Operators for PDEs on Complex Geometries and Topologies

37 views

Abstract

Learning PDE operators on complex domains requires capturing interactions among fields on vertices, edges, and faces, alongside global responses shaped by topology. Existing neural operators accommodate irregular geometries but often overlook these distinct field supports or their condition-dependent coupling. We introduce the , which combines topology-constrained interactions with learned coordination. DKHO encodes conditions on their native cochain supports, evolves Kuramoto-inspired relation states through the boundary and coboundary operators that compose the Dirac operator, and decodes non-harmonic and harmonic responses in orthogonal Hodge subspaces. Topology thus determines where information can flow, while learned dynamics adapts how it is exchanged to each PDE instance. Across porous-medium Darcy flow, torus transport–diffusion, and cavity magnetostatics, DKHO-large reduces prediction error % on average over leading baselines, while DKHO-small remains competitive % . These results suggest that coupling topological structure with adaptive dynamics provides an effective inductive bias for accurate and parameter-efficient PDE operator learning on complex geometries and topologies.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
neural operator, simplicial kuramoto dynamics, hodge decomposition, partial differential equations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Fe4SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes a Dynamic Kuramoto–Hodge Operator (DKHO) based on the three main components. The first component is a form-aware encoder, representing known PDE conditions on their native geometric support. The second component is Dynamic Simplicial Kuramoto module, coordinating along the adjacent-degree pathways of the Dirac operator. The third component is a Hodge-based decoder, decomposing the response into non-harmonic and harmonic responses. Overall, the paper addresses a largely underexplored problem, is clearly written and provides a new perspective for neural operators (NOs) on non-Euclidean domains. On the other hand, the paper misses some very important prior work, which somewhat undermines the novelty of the presented approach.

Critical strengths

The paper addresses a problem which is widely met and very important but is not yet well investigated, that is, neural PDE solvers on domains with complex geometries.

The paper is overall clean mathematically.

It provides quite a comprehensive list of the considered datasets with complex geometries.

Critical weaknesses

The paper misses some important, closely related work, thereby undermining the claims of novelty. See, for example, Global Topological Dirac Synchronization (Carletti et al,, 2025) and Dirac–Bianconi Graph Neural Networks – Enabling Non-Diffusive Long-Range Graph Predictions (Nauck et al., 2024). It seems that these results, especially, Dirac–Bianconi GNNs (DBGNNs) of Nauck et al., 2024 has a similar idea on Dirac-operator dynamics as a learned layer, while differs from DKHO on some other aspects. It is hence essential to discuss these approaches, contrasting to these results on the topological Dirac operator as well as the Dirac and global topological synchronization models, and present a numerical study comparing to DBGNNs. Otherwise, the novelty claim related to the Dirac–Kuramoto model is ungrounded.

Other missing competitors are Curv-GT of Liao et al., 2025 and GLNO of Tang et al. 2025.

A more extensive ablation study on the second component, i.e., Dynamic Simplicial Kuramoto module, will be helpful.

Results lack variability analysis, no standard deviations are reported.

Questions for authors

see weaknesses above

9 comments
AuthorsOpenReview ↗

Response to Reviewer Fe4S (Part 1)

We thank Reviewer Fe4S\color{green}{\mathbf{Reviewer~Fe4S}} for pointing out the missing related work, and for requesting a broader analysis of the Dynamic Simplicial Kuramoto module and experimental variability. These suggestions helped us sharpen both the positioning and the empirical evaluation.


1. Related work and novelty (W1; Q1)

We agree that Global Topological Dirac Synchronization and Dirac--Bianconi Graph Neural Networks (DBGNNs) are closely related and should be discussed explicitly. The revised discussion now distinguishes the questions addressed by these methods. Global Topological Dirac Synchronization studies when oscillators defined on different simplex degrees synchronize under a topological Dirac operator. DBGNNs learn non-diffusive node--edge propagation for long-range graph prediction. By contrast, DKHO shares their incidence-driven organization, but it is a condition-to-field neural operator: it separates persistent relation states from physical content states, adapts the exchange to PDE coefficients, forcing, and boundary conditions, and decodes fields directly on vertex, edge, or face supports through a Hodge-structured output map.

We also added the two missing competitive neural operators, CurvGT and GLNO, under both native-input and independently tuned matched-input protocols on all eight targets. Representative native-input results are:

MethodDarcy C1C^1 RelL2 ↓\downarrowTorus C1C^1 RelL2 ↓\downarrowCavity H0H^0 RelL2 ↓\downarrow
CurvGT0.08540.22710.0162
GLNO0.31040.50100.0190
DKHO-small0.02640.13100.0179
DKHO-large0.01550.12620.0140

DKHO-large outperforms both added competitors on all three representative targets. DKHO-small also improves substantially over both methods on Darcy C1C^1 and Torus C1C^1, while remaining close to CurvGT on Cavity H0H^0. The complete tables report all eight outputs under both native- and matched-input protocols, including the cases in which another method is stronger.

We agree that a numerical comparison with DBGNN would also be valuable. However, a faithful transfer is not a drop-in baseline: the public implementation predicts node--edge graph quantities, whereas our benchmarks map PDE conditions to vertex-, edge-, and face-supported fields. Extending DBGNN to this setting requires a rank-2 representation, condition encoder, face-supported output head, capacity matching, and independent tuning. A partial port would not provide a faithful comparison with the published DBGNN. Within the rebuttal period, we therefore focused the numerical study on CurvGT and GLNO, which could be adapted and validated consistently, while adding a direct conceptual comparison with both Dirac-based works.

We have expanded the Related Work to compare DKHO with Global Topological Dirac Synchronization and DBGNNs, and added the complete CurvGT and GLNO results to Sections 5.1--5.4 and Appendix E.1 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Fe4S (Part 2)

2. Dynamic Simplicial Kuramoto ablations (W2; Q2)

We agree that the original component study should have fully separated the individual choices inside the Dynamic Simplicial Kuramoto module. During the rebuttal, we therefore added both component removals and same-backbone replacements of the complete relation update. All variants use the same inputs, widths, data splits, and validation-based selection as DKHO-small:

Controlled changeMean task-wise RelL2 increase
Without Dirac routing+278%
Without phase rotation+27%
Without phase lag+22%
Without multiplicative gate+21%
Learned DEC diffusion+187%
Gated message passing+48.6%

The results separate the effects of routing and relation dynamics. Removing the signed Dirac routes has the largest average effect (+278%) and reaches +1098% on Darcy C2C^2. Face-level accuracy is therefore particularly sensitive to the cross-degree incidence route. Within the relation update, removing phase rotation, lag, or the multiplicative gate increases mean RelL2 by 27%, 22%, and 21%, respectively.

The two replacement controls test whether a more conventional update is sufficient. Degree-wise DEC diffusion preserves Hodge-Laplacian smoothing but removes the persistent cross-degree relation state, increasing error by 187% on average. Gated message passing preserves adaptive adjacent-rank exchange but replaces accumulated relation dynamics with direct content messages, increasing mean RelL2 by 48.6%. Together with ablations of component removal, these results show that the gains depend on both the signed incidence route and the way relation information is accumulated.

Representative ablations have been summarized in Section 5.5, and the complete eight-target accuracy and fidelity results, together with the two replacement dynamics, have been added in Appendix E.2 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Fe4S (Part 3)

3. Variability and standard deviations (W3; Q3)

We appreciate the request for multi-seed statistics. The complete comparison contains eight targets and ten models, including separately tuned native- and matched-input runs. Given the rebuttal time and compute budget, we completed the full comparison once with seed 42 rather than reporting multi-seed results for only a small subset. All reported runs use common data splits, evaluation samples, metric definitions, and validation-based model selection.

The main performance gaps are particularly large: DKHO-large obtains 0.0155 vs. 0.0854 on Darcy C1C^1 and 0.0436 vs. 0.6250 on cavity H2H^2, corresponding to approximately 82% and 93% lower RelL2 than the strongest native baselines. Together with the consistent protocol and the same ordering across the broader target set, these gaps increase the evidence of the advantage of our DKHO. We will run a broader multi-seed study in the camera-ready version.

We have clarified the common selection procedure, the evaluation samples, and the seed used for the reported DKHO runs in Appendices D.2--D.3 of the revised manuscript.

We thank the reviewer again for the constructive feedback, which has helped us position DKHO against the relevant competitors, examine the dynamics in isolation, and describe the stability of our results more candidly. We welcome further questions and discussion.

Reviewer_AsRUOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript propose a method that use Dynamic Kuramoto–Hodge Operator (DKHO) to learn PDE solution operators on domains, combined with complex geometries and topologies. The core idea of this paper is to combine discrete differential operators with Kuramoto-inspired dynamics, aiming to model interactions between fields defined on vertices, edges, and faces. Further more, The authors use a form-aware encoder to represent PDE conditions, a conditional Dirac–Kuramoto module to adapt information exchange across different cochain degrees. Also, the authors use a Hodge-based decoder to separate harmonic and non-harmonic responses.DKHO are evaluated on three PDE problems, which are arcy flow, transport–diffusion on a torus, and cavity magnetostatics. The proposed method is compared with six neural operator baselines, showing improvements in physical consistency and prediction accuracy. The ablation studies evaluate the contributions of the core components.

Critical strengths
  1. DKHO encodes discrete geometric topology into the neural computation flow. It enforces strict information propagation across difference geometric support and the condition-dependent parameters allowing flexible information exchange. This architectural design balances spatial constraints with adaptability, presenting a clear and targeted motivation
  2. Drawing on Hodge decomposition, the model orthogonalizes the output into non-harmonic and harmonic responses, providing a mathematically grounded and more interpretable division of representations.
  3. The experimental design is exceptionally thorough, effectively validating the core contributions of the paper.
Critical weaknesses

Although experiments show that DKHO models gained significant advantage over baselines even with less parameters (Table1&2), there is an unmatched training budget where DKHO models were trained for up to 300 epochs on several tasks (Table 10) while all baseline models were fixed at 200 epochs (Table12). Such setting makes it difficult to tell wheater the advantages were gained from extend training or architecture. Authors should either evaluate the DKHO and baselines under same training budget or at a mininum provide traning cureves to demonstrate wherther the baselines already fully converged at 200 epochs.

Questions for authors
  1. Why was unmatched training budget chosen for DKHO on several tasks?Could authors evaluate DKHO models on matched budgets to ensure a fair comparison?Alternatively, could authors provide training curves for baselines to demonstrate whether they had fully converged?
  2. The DKHO architecture elegantly reduce computation cost by limiting spatial exchange to low-dimensional phase cθ≪hc_\theta \ll h . Although the ablations have shown the necessity for spatial exchange, I am particularly interested in whether authors observed an accuracy bottleneck when cθc_\theta is set even lower, or would a slightly larger cθc_\theta brings any performance gains before computation cost becomes prohibitive?
2 comments
AuthorsOpenReview ↗

Response to Reviewer AsRU (Part 1)

We thank Reviewer AsRU\color{blue}{\mathbf{Reviewer~AsRU}} for recognizing the topology-aware design and for raising two important questions about training-budget fairness and the capacity of the low-dimensional relation state. We address them below.


1. Training-budget fairness (W1; Q1)

We agree that the different epoch budgets require a clear explanation. During the rebuttal, we re-trained the configurations affected by the earlier budget mismatch and made the schedules and validation-selected checkpoints consistent throughout the paper. We also added complete training curves for every method and target, following the alternative suggested by the reviewer.

The curves show that the baselines generally flatten within the first 50--100 epochs and change little thereafter. In contrast, both DKHO scales continue to improve during their longer task-specific schedules, and all 16/16 DKHO model--target checkpoints are selected within the final 10% of their budgets, indicating the necessity of the training protocol. The longer schedule is also coupled to the optimizer: the learning rate follows a cosine decay from 10−310^-3 to zero over the full budget. The additional epochs provide the low-learning-rate phase during which the DKHO validation loss continues to decrease. They are therefore part of the observed optimization behavior rather than repeated epochs after convergence.

This analysis also makes the trade-off explicit. DKHO uses more optimization steps on the harder supports, and we report the corresponding measured training time rather than treating the longer schedule as cost-free. The key fairness point is that extending the baseline schedules would not bring any performance improvements, whereas truncating DKHO would stop its schedule while validation error is still improving.

We have added the complete training curves and convergence discussion to Appendix D.4, clarified all budgets, optimizer schedules, and selected epochs in Appendix D.3, and reported measured training cost in Appendix C of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer AsRU (Part 2)

2. Phase-space dimension (Q2)

The quantity cθc_\theta denotes the channel dimension of the relation state, i.e., the dimension of the low-dimensional space in which spatial exchange is coordinated. Because the content width hh is held fixed, varying cθc_\theta isolates the capacity of the coordination mechanism. It is also the only design variable in this sweep that changes the parameter count. We therefore vary cθc_\theta while holding all other settings fixed.

cθc_\thetaDarcy C0C^0 RelL2 ↓\downarrowTorus C1C^1 RelL2 ↓\downarrow
20.004390.13224
40.004490.12778
60.004510.12862
80.004270.12848

A slightly larger phase space does help, but the improvement is not monotonic: Darcy C0C^0 peaks at cθ=8c_\theta=8 and Torus C1C^1 at 4, and the whole sweep spans 5.5% and 3.4% of RelL2 while the parameter count grows from 64.1K to 76.2K. Extra width therefore does not buy accuracy cheaply, which is a 19% increase in parameters moves the error by at most 5.5%. We choose cθ=2c_\theta=2 because it is the smallest point in the grid, and increasing the dimension does not bring obvious performance improvements. We therefore retain cθ=2c_\theta=2 for DKHO-small and use cθ=4c_\theta=4 for DKHO-large, matching the content width of the large model.

The sweep results have been added in Section 5.6, and the small/large relation widths, parameter growth, and capacity trade-off have been discussed in Appendix C of the revised manuscript.

We thank the reviewer again for helping us clarify how the training budget is accounted for and how the capacity of the relation state was chosen, which we believe makes our empirical claims considerably easier to interpret. Further discussion is more than welcome.

Reviewer_uk6XOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes the Dynamic Kuramoto–Hodge Operator (DKHO), a neural operator for learning PDE solutions on domains with complex geometries and nontrivial topologies. The method combines form-aware encoding of fields on vertices, edges, and faces, Kuramoto-inspired dynamics for adaptive information exchange, and Hodge-structured decoding to separate harmonic and non-harmonic responses. The authors evaluate DKHO on Darcy flow, torus transport–diffusion, and cavity magnetostatics. The results show consistent improvements over several neural operator baselines, while the smaller DKHO variant achieves competitive accuracy with substantially fewer parameters. Overall, I find the approach interesting and the results promising, although I have some concerns about the specific benefits of the Kuramoto dynamics, the interpretation of the harmonic responses, and the experimental validation.

Critical strengths
  • Well-motivated architecture: The paper addresses an important challenge in neural operator learning: representing physical fields on different geometric supports while accounting for domain topology. I find the combination of discrete exterior calculus, adaptive relation dynamics, and Hodge decomposition well motivated.

  • Interesting use of Kuramoto dynamics: The separation between content states and relation states is an interesting design choice. It allows the model to adapt information exchange without directly treating the oscillator phases as physical quantities. The mathematical formulation in Section 4 is also reasonably clear.

  • Strong experimental results: DKHO-large achieves the lowest reported relative L2 error across all seven prediction targets. The improvements also extend to differential-fidelity and structural metrics. I particularly appreciate the matched-condition comparisons in Appendix E.1, which help evaluate whether the improvements come from the architecture or the additional input information.

  • Good parameter efficiency: DKHO-small uses approximately 64K parameters, substantially fewer than the main baselines, while remaining competitive across the evaluated tasks. I think this is a useful result, especially for applications where model size is important.

  • Useful ablation studies and technical details: The ablations in Table 21 provide evidence for the contributions of Dirac routing, relation dynamics, and harmonic decoding. The appendices also include mathematical explanations and implementation details that help me understand the proposed method.

Critical weaknesses
  • The specific benefit of Kuramoto dynamics is not fully established: Although the proposed combination is interesting, several underlying ideas have already appeared in related work. Topological Neural Operators (TNO) [1] use discrete exterior calculus for cross-dimensional interactions, Hodge Spectral Duality (HSD) [2] separates harmonic and non-harmonic responses, and the Kuramoto Neural Operator (KNO) [3] explores oscillator dynamics for PDE operator learning. The authors discuss these works, and I recognize that DKHO combines these ideas in a different way. However, I am not fully convinced that the experiments isolate the advantage of the specific Kuramoto-inspired dynamics over simpler adaptive cross-rank interactions. The no-phase ablation is useful, but removing the relation state entirely is different from replacing its dynamics with a simpler alternative. I think a controlled comparison with a parameter-matched DEC-based model using conventional learned gates or nonlinear diffusion would make the architectural contribution clearer.

  • The cavity experiment does not directly validate the H² harmonic interpretation: I find the cavity magnetostatics experiment interesting, but I have a concern about its topological interpretation. The spherical shell has nontrivial second cohomology, with β₂ = 1, which is relevant to the cavity-related magnetic flux. However, Appendix D.3 states that the implemented cavity output head uses a nodal H⁰ construction rather than a face-supported H² harmonic projection. The authors are clear about this distinction, but constant nodal modes also exist on simply connected domains. Therefore, I am not fully convinced that the reported results directly demonstrate the benefit of learning the cavity-specific H² response. An evaluation of face-integrated fluxes or a comparison with a simply connected domain would help clarify this point.

  • Baseline fairness and the role of input features need further clarification: I appreciate that the authors include matched-condition comparisons in Appendix E.1. However, these results also show that the additional input descriptors account for part of the original performance gap. Based on the reported relative L2 errors, I calculate an average improvement of approximately 56.6% over the strongest native-input baselines, compared with 32.2% under matched-input comparisons. The remaining improvement is still meaningful, but I think it is important to better separate the effects of the architecture and the input features. I also noticed that some baselines become substantially worse after input augmentation. For example, HSD's torus concentration RelL2 increases from 0.3613 to 1.6959, while its cavity RelL2 increases from 0.0547 to 0.2156. I would like to know whether the augmented baselines were independently retuned and why these results deteriorate so much.

  • Limited generalization and additional validation: All three benchmarks use fixed meshes and mainly evaluate generalization to new input conditions. For example, the Darcy problem uses a fixed permeability field, while the torus problem fixes the geometry, diffusivity, velocity configuration, and terminal time. This makes it difficult to judge how well DKHO transfers to different resolutions or physical configurations. I understand that generalization to unseen geometries is not directly claimed as an experimental result, but I think a small remeshing or parameter-variation experiment would strengthen the broader motivation. Also, since the output fields are trained separately, good individual prediction accuracy does not necessarily guarantee consistency between related physical quantities. Additional PDE residual or conservation metrics would be useful. Finally, multi-seed statistics and measured inference costs would help assess the reliability and practical efficiency of the reported improvements.

[1] Bastian, L., Leventhal, S., Hajij, M., and Birdal, T. (2026). Topological Neural Operators. arXiv:2606.09806. https://arxiv.org/abs/2606.09806

[2] Zheng, D., Zhong, T., and Allen-Blanchette, C. (2026). Topology-Preserving Neural Operator Learning via Hodge Decomposition. Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306, pp. 164029–164060. https://proceedings.mlr.press/v306/zheng26aj.html

[3] Badolia, P., Obukhov, L., Bylinkin, D., and Beznosikov, A. (2026). The Kuramoto Neural Operator: Learning to Solve PDEs via Coupled Oscillator Dynamics. arXiv:2608.10234. https://arxiv.org/abs/2608.10234

Questions for authors
  1. What is the specific advantage of the Kuramoto dynamics over simpler adaptive interactions? Could the authors compare DKHO with a parameter-matched model that keeps the same Dirac routing and input features but replaces the Kuramoto evolution with conventional learned diffusion or gated message passing? I think this is important for understanding the main architectural contribution.

  2. How sensitive is DKHO to the relation dynamics? In particular, how do the number of phase substeps, learned gates, and phase lags affect prediction accuracy? Would a simpler relation update achieve similar results?

  3. How does the nodal H⁰ decoder demonstrate the benefit of the cavity's H² topology? Appendix D.3 explains the implementation, but I would appreciate more discussion of how the nodal harmonic representation relates to the cavity-specific global flux response.

  4. Can the harmonic contribution in the cavity experiment be validated more directly? For example, could the authors evaluate face-integrated magnetic fluxes using an H² basis or compare the spherical shell with a simply connected domain?

  5. Were the matched-condition baselines independently retuned after input augmentation? I noticed that some HSD results deteriorate considerably under the matched-input protocol. Could the authors explain these differences and clarify the optimization procedure?

  6. How much of DKHO's performance improvement comes from the additional input descriptors? Would it be possible to report a focused ablation removing the spectral positional encodings, input diffusion descriptors, or global summaries while keeping the rest of DKHO unchanged?

  7. How well does DKHO generalize to changes in mesh resolution or physical conditions? For example, could the authors provide a small cross-resolution or permeability-variation experiment? I think this would help clarify the practical scope of the method without requiring a new benchmark.

  8. Can the authors provide additional evidence of physical consistency and experimental reliability? Since the fields are predicted separately, I would be interested in their conservation or constitutive consistency. If available, multi-seed variability and measured inference-time results would also help interpret the reported accuracy and parameter-efficiency gains.

10 comments
AuthorsOpenReview ↗

Response to Reviewer uk6X (Part 1)

We thank Reviewer uk6X\color{red}{\mathbf{Reviewer \ uk6X}} for the thoughtful and detailed feedback. The review raises four important issues: isolating the contribution of the Kuramoto relation dynamics, directly validating the cavity-specific H2H^2 response, separating architectural gains from input-feature gains, and assessing transfer, physical consistency, and experimental reliability. We address these points in order below.


1. Kuramoto dynamics: controlled alternatives and sensitivity (W1; Q1--Q2)

We agree that the no-phase ablation alone does not isolate the relation dynamics. During the rebuttal, we additionally performed ablations of varying the phase substeps, removing the lag/gate, and replacing the Kuramoto update:

Controlled settingDarcy C0C^0 RelL2 ↓\downarrowTorus C1C^1 RelL2 ↓\downarrowMean RelL2 change over 8 targets
Substeps R=2R=20.004390.13224--
Substeps R=4R=40.003940.13034--
Substeps R=6R=60.003800.13059--
Substeps R=8R=80.003980.12963--
Without phase lag0.00470.1360+22%
Without learned gate0.00460.1366+21%
Decoupled DEC diffusion0.01340.1490+187%
Gated message passing0.00640.1749+48.6%

Three patterns are clear from these results. First, performance changes only modestly over R=2R=2--88: Darcy C0C^0 improves from 0.00439 at R=2R=2 to 0.00380 at R=6R=6, while Torus C1C^1 varies within the narrow range 0.12963--0.13224. We therefore use R=2R=2 as an efficient default rather than relying on a finely tuned number of substeps.

Second, the gate and lag make distinct, measurable contributions. The gate controls how strongly an adjacent-degree relation state enters the update, whereas the lag specifies the preferred offset of the nonlinear response. Removing either component increases mean RelL2 by 21--22% across the eight targets.

Third, the simpler replacements do not recover the same overall performance. Decoupled DEC diffusion relaxes each degree independently under its Hodge Laplacian and does not maintain a persistent cross-degree relation state. Gated message passing exchanges adjacent-rank content directly, but does not accumulate relation information over multiple substeps. Their mean RelL2 increases by 187% and 48.6%, respectively. These controls support the value of accumulating condition-dependent coordination along the signed incidence routes, beyond simply adding another adaptive message-passing block.

We have added the substep sensitivity analysis to Section 5.6, the representative gate/lag and replacement results to Section 5.5, and the complete eight-target comparison to Appendix E.2 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer uk6X (Part 2)

2. Direct validation of the cavity H2H^2 response (W2; Q3--Q4)

The spherical shell has β2=1\beta_2=1, with a unique global flux mode represented on 2-cochains. The original readout, however, was a nodal vector field. A nodal constant mode exists on any connected domain, which means the original evaluation could not certify that DKHO had captured anything cavity-specific. The relation between the two modes is an integral mode. To obtain a flux from a nodal field at all, one has to aggregate it over a closed surface, ΦS(h)=∑f∈S∣f∣ nf⋅avg(h)\Phi_S(h)=\sum_{f\in S}|f|\,n_f\cdot\mathrm{avg}(h), whereas the cavity mode itself is the β2=1\beta_2=1 element of ker⁡Δ2\ker\Delta_2 — a closed but non-exact face flux that can be represented at degree two but not carried by vertex values, and no term in the nodal objective supervises that aggregation.

To address this concern directly, during the rebuttal we adopted the first validation proposed by the reviewer and added a separately trained, face-supported target. The new decoder reads rank-2 content on oriented faces and predicts the face-integrated magnetic flux as a non-harmonic residual plus a coefficient in a supplied combinatorial H2H^2 basis. For the spherical shell with β2=1\beta_2=1, this single coefficient measures the amplitude of its unique cavity-enclosed global-flux mode.

Method (matched inputs)H2H^2 RelL2 ↓\downarrowH2H^2 coeff. MAE ↓\downarrowDiv Fid ↑\uparrowIoU ↑\uparrow
GNO0.55300.74100.04490.8075
DKHO-small0.05260.10850.68750.9500
DKHO-large0.04360.10930.77910.9577

The four columns score the face field itself: RelL2 over the whole flux, coefficient MAE for the amplitude of the global mode, divergence fidelity for its differential consistency, and IoU for where the strong response sits. Under matched inputs, DKHO-large reduces RelL2 from 0.5530 to 0.0436. Both DKHO variants bring the coefficient error down to approximately 0.109, and divergence fidelity and overlap rise from 0.0449 and 0.8075 to 0.7791 and 0.9577. Since these errors are produced by a model whose output space is the face cochain, they directly measure the flux mode rather than through a post-hoc integration of vertex values, which we hope address the concerns.

We have introduced the face-supported H2H^2 task and its results in Section 5.4, documented the data construction and decoder in Appendices D.1 and D.3, and provided the full native- and matched-input metrics in Appendix E.1 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer uk6X (Part 3)

3. Baseline fairness and feature attribution (W3; Q5--Q6)

Independent re-tuning. Yes. Every baseline was tuned independently under the native- and matched-input protocols. For each protocol, we screened a 3×33\times3 grid over learning rate and weight decay for 10 epochs, refined the best cell with three neighboring values, selected the checkpoint on the common validation split, and evaluated the held-out test split once. The matched HSD numbers are therefore obtained from separately tuned runs, not from native checkpoints with additional channels appended afterward.

The HSD deterioration is a genuine property of that architecture in this setting rather than an artefact of stale hyperparameters. Its spectral branch already carries global spectral and harmonic information. The added Laplacian positional encodings, heat descriptors, and global summaries are thus partly redundant and alter the input scaling seen by its spectral decoder.

Even after independent re-tuning, the validation-selected matched configuration further deteriorates on Torus C0C^0 (0.3613 →\rightarrow 1.6959) and Cavity H0H^0 (0.0547 →\rightarrow 0.2156).

HSD RelL2 ↓\downarrowTorus C0C^0Cavity H0H^0
Native inputs0.36130.0547
Matched inputs1.69590.2156
Change+369%+294%

Feature attribution. To quantify how much DKHO benefits from the additional descriptors, we additionally removed one feature family at a time while keeping the DKHO-small backbone, widths, splits, and training protocol fixed. Representative results across the three benchmarks are:

DKHO-small inputDarcy C1C^1 RelL2 ↓\downarrowTorus C1C^1 RelL2 ↓\downarrowCavity H0H^0 RelL2 ↓\downarrow
Full input0.02640.13100.0179
Without Laplacian positional encoding0.0363 (+38%)0.1338 (+2.1%)0.0211 (+18%)
Without heat descriptors0.0348 (+32%)0.1337 (+2.1%)0.0215 (+20%)
Without global summary0.2438 (+824%)0.1356 (+3.5%)0.2086 (+1065%)

The above results well correspond to where the information actually lives. The global summary dominates wherever the forcing is localized (Darcy and the cavity). It is also almost free on the uniform torus, whose excitation is a smooth blob that the local descriptors already place (+3.5%). The positional and heat descriptors contribute a steady 12–16% on average. The richer condition description therefore explains part of the original gain, but not all of it: when every method receives the same descriptors, DKHO-small still attains the lowest RelL2 in 6/8 matched comparisons, compared with 5/8 under native inputs. Together, the two extra controls separate the contribution of the inputs from that of the architecture.

The re-tuning procedure is documented in Appendix D.2, the complete native- and matched-input comparisons in Appendix E.1, and the full eight-target feature ablations in Appendix E.2 of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

To All Reviewers: Updated Manuscript and Appendix

We thank all reviewers for their careful and constructive feedback. In addition to the individual responses below, we have revised the manuscript and appendix to address the questions and suggestions raised. All revisions are marked in red. The main changes are summarized below:

  1. @Reviewer uk6X\color{red}{\mathbf{@Reviewer\ uk6X}} and @Reviewer Fe4S\color{green}{\mathbf{@Reviewer\ Fe4S}}: Controlled relation-dynamics comparisons. We added same-backbone alternatives that replace the Kuramoto update with degree-wise DEC diffusion or gated adjacent-rank message passing, together with component ablations of Dirac routing, phase rotation, phase lag, the multiplicative gate, and the harmonic branch in Section 5.5 and Appendix E.2.
  2. @Reviewer uk6X\color{red}{\mathbf{@Reviewer \ uk6X}}: Direct cavity-topology evaluation. We added a face-supported H2H^2 target of oriented area-integrated magnetic flux and evaluated its harmonic-basis coefficient, boundary fluxes, conservation, divergence fidelity, and spatial overlap in Section 5.4 and Appendices D.1, D.3, and E.1.
  3. @Reviewer uk6X\color{red}{\mathbf{@Reviewer \ uk6X}}: Baseline fairness and feature attribution. We independently tuned every matched-input baseline and added fixed-backbone ablations of Laplacian positional encodings, heat descriptors, and global summaries in Appendices D.2, E.1, and E.2.
  4. @Reviewer uk6X\color{red}{\mathbf{@Reviewer \ uk6X}} and @Reviewer Fe4S\color{green}{\mathbf{@Reviewer \ Fe4S}}: Generalization, physical consistency, reliability, and cost. We added zero-shot permeability-contrast and four-fold mesh-refinement tests, additional physical and structural metrics, a unified seed-42 protocol, and measured end-to-end inference and training costs in Section 5.6 and Appendix C.
  5. @Reviewer AsRU\color{blue}{\mathbf{@Reviewer \ AsRU}}: Training dynamics and phase-space dimension. We re-ran configurations affected by the earlier budget mismatch, unified the reporting of schedules and selected checkpoints, added training curves for every method and target, and swept the phase width cθc_\theta in Appendices D.3–D.4 and Section 5.6.
  6. @Reviewer Fe4S\color{green}{\mathbf{@Reviewer \ Fe4S}}: Related work and competitive baselines. We clarified the relationship between DKHO and Global Topological Dirac Synchronization, and DBGNNs, and added CurvGT and GLNO under native- and matched-input protocols on all eight targets in the Related Work, Sections 5.1–5.4, and Appendix E.1.

We appreciate the reviewers' suggestions, which have helped us clarify DKHO's architectural contribution, topological interpretation, experimental fairness, and practical cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.