acceptodds
Under review as a conference paper at ICLR 2027

Where to Compute and How to Interact: Operator-Readable Adaptation with Gauge-Aware Transport

Abstract

Adaptive meshes enable neural operators for partial differential equations (PDEs) to allocate spatial samples and computational resources according to local physical structures. Existing approaches, however, primarily focus on where to compute, while paying less attention to how to interact after node relocation. Mesh adaptation changes local sampling scales, neighborhood structures, and geometric contexts, making representations formed at different nodes not necessarily directly comparable. Direct aggregation may therefore entangle genuine physical variation with discretization-induced representation variation. Moreover, because allocation and interaction are jointly optimized through the same output objective, their individual roles are difficult to distinguish from final errors alone. We introduce operator readability, which requires an adaptive operator to explicitly account for and test why computation is allocated to particular locations and how representations interact under the resulting nonuniform discretization. Based on this principle, we propose the Gauge-Aware Adaptive Mesh Neural Operator (GA-AMNO). Physics-informed adaptive allocation answers where to compute, while geometry-conditioned low-rank Gauge transport maps source features into target representation contexts before aggregation, answering how to interact. This design turns the otherwise implicit mesh-to-solver information exchange into an inspectable and intervenable computational process. We further establish sufficient conditions for representation-consistent aggregation and analyze approximate transport errors and continuity under topology-preserving mesh deformations. Experiments across five PDE benchmarks demonstrate improved predictive accuracy, while controlled interventions and geometric-mismatch analyses verify the computational roles of allocation and interaction and show that Gauge transport improves cross-discretization representation compatibility under strong geometric mismatch.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, adaptive meshes, gauge-aware representation, operator readability, partial differential equations, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Pue4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a framework designed to address two interconnected aspects of adaptive neural operator learning.

The method combines physics-informed adaptive node allocation with geometry-conditioned low-rank feature transport. The allocation module uses physical indicators, including gradients, Laplacians, energy, and vorticity-related quantities, to guide node relocation. The transport module maps source-node features into target-conditioned representation contexts before aggregation. The architecture additionally incorporates spectral and differential residual corrections.

The authors introduce the concept of operator readability and provide theoretical conditions for representation-consistent aggregation and approximate transport. Experiments on five PDE benchmarks: Navier–Stokes, Rayleigh–Bénard convection, Kuramoto–Sivashinsky, Kolmogorov flow, and Darcy flow, demonstrate competitive prediction accuracy. Ablation studies, allocation counterfactuals, edge interventions, and geometric-mismatch analyses further investigate the behavior of the proposed mechanisms.

Critical strengths
  1. A formulation of adaptive computation that considers both resource allocation and feature interaction.
  2. Experimental evaluation with informative mechanism-level interventions.
Critical weaknesses
  1. The benefit specifically attributable to Gauge transport is not fully isolated from architectural capacity and other correction mechanisms.
  2. The theoretical interpretation and generalization of the learned transport remain limited.
Questions for authors
  1. compare GA-AMNO with a parameter and compute matched geometry conditioned message passing architecture that does not use the proposed Gauge transport parameterization. Both models should retain the same adaptive allocation, spectral correction, and differential correction modules, with independently trained parameters. This would help isolate the specific contribution of Gauge transport.
  2. How does GA-AMNO perform on substantially nonuniform or anisotropic meshes, or under larger discretization changes than those evaluated?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Pue4 (1/2)

W1 & Q1. Isolating the benefit of Gauge transport from capacity and correction mechanisms.

A1. We use ablations to clarify what each component contributes. Allocation shapes interaction geometry, Gauge transport transforms messages, and the correction modules refine the output field. Classic serves only as a system-level comparison between the full GA-AMNO architecture and a conventional adaptive solver. In Fig. 4, Identity Transport replaces learned transport with the identity map while retaining the adaptive mesh, importance weighting, and both spectral and differential corrections. This replacement increases prediction error, most notably on Kolmogorov. The allocation-module ablation in Fig. 4 fixes the node coordinates, switches off importance-based feature gating and the physical-allocation loss, and retains gauge transport. The separate importance-map interventions distinguish the displacement and feature-gating paths, and Fig. 8 reports targeted ablations of the correction modules.

We compare GA-AMNO with an independently trained dense-kernel control. With the other modules unchanged and comparable capacity and compute, GA-AMNO achieves better spatial accuracy on all three benchmarks. The control replaces only the Gauge matrix-generation heads with an MLP that generates a full transport map from the same edge descriptor. Scalar attention, adaptive allocation, mesh-to-grid reconstruction, and the spectral and differential prediction heads are retained. Both variants are trained independently with comparable parameter counts and computational costs. Appendix C.4.3 specifies the protocol, and Appendix F.2, Table 5 reports the results below.

DatasetMean relative ℓ₂: OursMean relative ℓ₂: MLP kernelHigh-gradient region: OursHigh-gradient region: MLP kernelLaplacian: OursLaplacian: MLP kernelSpectral: OursSpectral: MLP kernel
Navier–Stokes0.0031520.0035730.0061520.0072440.0320080.0376620.0014950.001470
Kolmogorov0.0186120.0196650.0240270.0253710.1149230.1166670.0064030.008073
Darcy0.0268270.0284500.0228850.0249630.4350250.4479530.0213520.021964

The matched comparison shows that the spatial gains extend to regions where the solution varies sharply. Edge interventions further show that replacing transport on top-ranked edges changes predictions more than replacing it on random or bottom-ranked edges. With the same prediction heads, structured transport achieves lower full-field, high-gradient-region, and Laplacian errors across all three benchmarks. These gains extend beyond average field accuracy: the model better resolves regions where the solution varies sharply and more accurately captures second-order spatial variation. The dense kernel achieves slightly lower spectral error on Navier–Stokes, a result consistent with the spectral metric measuring Fourier energy without directly assessing spatial phase. Fig. 5(b) shows where the learned transformations matter most: replacing transport with the identity on top-ranked edges changes predictions more than doing so on random or bottom-ranked edges. Together, the matched comparison shows what the structured parameterization contributes to spatial accuracy, while the edge interventions identify which learned transformations have the greatest influence on the prediction.

AuthorsOpenReview ↗

Response to Reviewer Pue4 (2/2)

W2 & Q2. Theoretical interpretation and applicability under nonuniform discretizations.

A2. Our theory explains when source-to-target transport makes aggregation consistent across local representations, and what controls the error when the learned transport is only approximate. The key idea is to re-express each source feature in the target node’s representation before combining messages (Appendix D.2, Eqs. (19)–(20)). Theorem 1 concerns shared linear aggregation with unchanged scalar weights. Theorem 2 gives sufficient conditions for consistent aggregation: the transport map transforms consistently at both endpoints, and scalar attention weights stay unchanged. Theorem 3 then bounds the aggregation inconsistency by separating changes in neighbor weighting from errors in the channel transformation. Proposition 1 characterizes what the diagonal-plus-low-rank map can approximate, while Proposition 2 and Corollary 1 show how the effects of topology-preserving mesh deformations propagate from individual messages to the prediction under the stated assumptions. Exact equivariance is the reference case for changes of representation, not the objective under re-discretization. A change of frame re-expresses the same information, whereas a change in sampling scale, orientation, or spatial support can change the information the encoder receives. GA-AMNO therefore learns how to adapt each incoming message to the target’s local geometry before aggregation; the theory explains what this design buys us and how deviations from the ideal affect the aggregate.

We add a Darcy evaluation across nonuniform and directionally varying meshes. Section 4.4 and Table 2 compare regular, clustered, sheared, twisted, and adaptive layouts. Clustering changes node spacing and spatial distribution, while shear and twist alter local directions and neighborhood geometry. A model is trained separately for each prescribed mesh, and the identity intervention replaces transport in its selected checkpoint. The comparison therefore examines the predictive role of the same transport design across trained configurations.

MeshIdentity interventionGauge transportRelative error reduction
Regular0.0364620.02793123.40%
Clustered0.0848560.02848166.44%
Sheared0.1416530.02622981.48%
Twisted0.2339450.02660088.63%
Adaptive0.0570720.02682752.99%

Transport is useful across the tested geometries, with larger relative intervention effects on the nonuniform layouts. Errors with gauge transport lie within 0.026229–0.028481, compared with the wider error range under identity replacement. Its regular-grid benefit shows that explicit adaptation is not required, while the larger effects on clustered, sheared, and twisted meshes support considering local geometric context in information exchange. These results demonstrate the applicability of the proposed transport design across the tested mesh-specific configurations.


Summary Response

We thank you for recognizing the allocation–interaction formulation and mechanism-level evaluations. We add an independently trained dense-kernel control with the remaining modules retained to isolate the benefit of structured transport (A1). We clarify the assumptions and limits of the theoretical interpretation and evaluate nonuniform and directionally varying mesh configurations (A2). We hope these results and clarifications address your concerns about attribution and scope.

Reviewer_Y9viOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

GA-AMNO moves the nodes of a coarse grid, with displacements scaled by an importance map predicted from physics indicators. On the moved mesh, each edge of a fixed k-neighbor stencil gets a scalar weight α_ij and a diagonal-plus-low-rank transport T_{i <- j}. Both are predicted from an edge descriptor (relative position, mesh Jacobians, node state descriptors). GA-AMNO reports the lowest relative L2 error on five PDE benchmarks.

Critical strengths
  • Separating allocation from interaction is a relatively clean framing
Critical weaknesses
  • alpha_ij · T_ij is a constrained factorization of the edge-conditioned C×C kernel(e_ij) that Graph NO and Transolver already uses.
Questions for authors
  • How about we control for example extra inputs and other components the same across models? Did the baselines receive the same input channels and the same gradient, Laplacian, and spectral losses?

  • Does node density concentrate where those physical vars are large? I wonder whether the relocation acts as learned conditioning or resolution allocation? Node relocation does not change which input values a node samples, its neighbours, or where its output is decoded. It only changes the geometric part of e_ij from my understandings.

  • What is the contribution of the adaptive mesh itself (no feature alignment)? Or feature alignment with no adaptive mesh? Do the two coupled used?

  • A broader discussion with the authors: if there is no explicit "adaptive meshing" design, does the "feature alignment" design still need to be considered in cases like no features, concatenating features without neighbor info in the input, such as reference grid [1], Fourier Features [2], or downstream efforts on the embeddings like in [3] on irregular geometries? If so, I think this design can be broadly used in general neural operator methods; If not, setting a working boundary for this design on adaptive meshing is also okay.


[1] Wu, H., Luo, H., Wang, H., Wang, J., & Long, M. (2024). Transolver: A fast transformer solver for pdes on general geometries. arXiv preprint arXiv:2402.02366.

[2] Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., ... & Ng, R. (2020). Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33, 7537-7547.

[3] Song, H., Li, Z., Zhang, X., Gan, Z., Lai, Z., & Wang, W. (2025). Reformulating Neural Operators in $ d+ 1$ Dimensions for Embedding Evolution. arXiv preprint arXiv:2505.11766.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Y9vi (1/3)

W1 & Q1. Distinctive interaction design and controlled comparisons of inputs, architecture, and supervision.

A1. GA-AMNO uses the same physical inputs as the baselines and separates channel transformation from neighbor weighting, so the transformation can be tested on its own. Graph NO [1] uses matrix-valued integral kernels conditioned on spatial coordinates and input-field values to model local interactions. GA-AMNO differs in how it constructs and applies the kernel: relative coordinates after relocation, mesh Jacobians at both endpoints, and state descriptors determine a directed transport map with diagonal scaling and a gated low-rank correction. A separate scalar attention weight determines how much the transformed message contributes to aggregation (Eqs. (7)–(9)). Transolver [2] organizes spatial points into learned slices, applies attention across physics-aware tokens, and maps the updated tokens back to the points. Its slices organize information across spatial locations, whereas GA-AMNO’s low-rank correction mixes feature channels within each directed source-to-target interaction. This separation lets us test the channel transformation independently of neighbor weighting: in Fig. 5(b), we intervene on selected transport maps while holding mesh geometry and attention fixed. Section 3.4, “Relation to conditional operator kernels,” explains these distinctions in detail.

We train a separate control that uses an MLP to generate dense transport maps. GA-AMNO gives lower spatial errors on all three benchmarks. An MLP directly generates a full transport map from the same edge descriptor, replacing only the Gauge matrix-generation heads. Attention, adaptive allocation, reconstruction, and prediction heads are retained. The variants are trained independently with comparable parameter counts and computational costs (Appendix C.4.3; Appendix F.2, Table 5). Structured transport yields lower full-field, high-gradient-region, and Laplacian errors on all three tested benchmarks, while the dense kernel has slightly lower spectral error on Navier–Stokes. Because both variants learn edge-dependent matrices, this comparison tests the parameterization rather than merely the presence of a feature-transformation module.

DatasetMean relative ℓ₂: OursMean relative ℓ₂: MLP kernelHigh-gradient region: OursHigh-gradient region: MLP kernelLaplacian: OursLaplacian: MLP kernelSpectral: OursSpectral: MLP kernel
Navier–Stokes0.0031520.0035730.0061520.0072440.0320080.0376620.0014950.001470
Kolmogorov0.0186120.0196650.0240270.0253710.1149230.1166670.0064030.008073
Darcy0.0268270.0284500.0228850.0249630.4350250.4479530.0213520.021964

We also match output supervision to examine whether the auxiliary losses explain GA-AMNO’s predictive advantage. The original full-solver comparison retains each baseline’s architecture but does not use identical gradient, Laplacian, and spectral losses and coefficients across methods. We therefore retrain all seven baselines on Navier–Stokes using field MSE and the same auxiliary loss definitions and coefficients as GA-AMNO, while retaining each baseline’s applicable architecture-specific regularizers. The table reports mean per-sample relative ℓ₂ error on the same test split, with seed 42 for all methods. Several baselines, including FNO, improve under matched supervision, showing that the auxiliary losses contribute to prediction accuracy. With the same output supervision, GA-AMNO still achieves the lowest error: 0.003152, approximately 12.9% below FNO, the strongest baseline in this comparison. This result shows that GA-AMNO’s predictive advantage persists after matching the auxiliary losses.

MethodGA-AMNOClassicMMPDETransolverFNOGNOWNOUNet
Relative ℓ₂ ↓0.0031520.0101000.0042230.0185580.0036180.0307210.0157580.005805
AuthorsOpenReview ↗

Response to Reviewer Y9vi (2/3)

Q2. Node concentration and the role of relocation as geometric conditioning.

A2. We do not observe a pronounced, systematic increase in node density in high-importance regions. The importance map scales each node’s displacement magnitude without requiring neighboring nodes to converge on the same region (Eq. (6)). It also weights the updated nodal features before reconstruction (Eq. (22)). Consequently, a region can contribute more strongly through its existing features without gaining additional physical samples. The allocation-module and high-demand-region ablations in Figs. 4 and 8(a) evaluate the combined module, rather than establish density concentration.

Relocation lets the operator adapt how nodes exchange features to the input while preserving their sampled inputs, neighbors, and decoding sites. Each node retains its reference-grid input features, indexed neighbors, and output decoding site (Appendices D.2–D.3). Relocation changes relative positions and local Jacobians, which condition attention weights and source-to-target channel transformations (Eqs. (7)–(10)). It can therefore change how a node sends and receives information even when the sampled input and neighbor identities remain fixed. The distinction is that importance gating weights the final features used for reconstruction, while relocation changes the geometry used to exchange them beforehand. Making these pathways explicit and open to intervention is the operational meaning of readability in this implementation.


Q3. Separate contributions of allocation and transport and their computational coupling.

A3. The separate ablations in Fig. 4 highlight the contribution of each module: prediction is more accurate when adaptive allocation and learned transport are both retained. In Fig. 4, replacing gauge transport with the identity increases prediction error while retaining adaptive allocation, with the largest increase on Kolmogorov. Ablating the allocation module also increases error: this variant fixes node coordinates and disables importance-based feature gating and the physical-allocation loss while retaining learned transport. The separate interventions in Fig. 5(a) clarify the roles of the two allocation pathways. Feature gating accounts for nearly all of the joint error response, while the displacement-only changes produce little change in prediction error. This contrast reflects where the interventions act: the gate directly reweights the updated features that enter reconstruction, whereas displacement changes the geometry used to exchange messages upstream.

Allocation and transport work together through the geometry used to construct each incoming message. Allocation changes the relative positions and local Jacobians in the edge descriptor (Eq. (7)). Transport uses that descriptor to determine which combination of source features is sent (Eq. (8)), while attention determines how much the message contributes relative to other neighbors (Eq. (9)). This gives node relocation a concrete role even when the input samples, neighbor identities, and decoding sites remain fixed: moving a node changes how its features are exchanged. The Darcy results in A4 illustrate why this connection matters. Learned transport improves prediction on the regular grid, and its relative benefit is larger on every tested nonuniform mesh, especially the sheared and twisted layouts. The pattern is consistent with the design rationale: adjustment before aggregation is more valuable when neighboring nodes have more heterogeneous local sampling geometries.

AuthorsOpenReview ↗

Response to Reviewer Y9vi (3/3)

Q4. Applicability beyond explicitly adaptive meshes.

A4. GA-AMNO improves prediction even on a fixed grid, and its relative benefit is larger on every tested nonuniform mesh. Node relocation is not a prerequisite: with fixed coordinates, the state descriptors still vary with the input, allowing the source-to-target transport maps to adapt to the physical state. To test this across different geometries, we add a Darcy comparison in Section 4.4 and Table 2. For each mesh configuration, we compare gauge transport with identity replacement within the corresponding checkpoint.

MeshIdentity interventionGauge transportRelative error reduction
Regular0.0364620.02793123.40%
Clustered0.0848560.02848166.44%
Sheared0.1416530.02622981.48%
Twisted0.2339450.02660088.63%
Adaptive0.0570720.02682752.99%

With Gauge transport, prediction errors are comparable across all five Darcy meshes, and the gains are larger on the tested nonuniform meshes. This makes the source-to-target design useful on fixed grids and especially valuable when local sampling geometries differ. The sheared and twisted meshes benefit most: their errors are the highest under identity replacement, yet with Gauge transport they are comparable to the regular-grid result. This pattern supports the design rationale: when neighboring nodes have more heterogeneous sampling contexts, adjusting each incoming message before aggregation becomes more valuable. The part of GA-AMNO that can carry over to other operators is this source-to-target transformation. An operator with fixed reference coordinates or positional embeddings can combine them with latent states to condition its transport maps before aggregation, while retaining its existing grid and prediction architecture.


Summary Response

We thank you for recognizing the clear allocation–interaction framing. We clarify the relationship to conditional kernels and provide matched dense-kernel and output-supervision comparisons (A1). We explain relocation as geometric conditioning rather than demonstrated density concentration (A2), distinguish the module contributions and computational dependency (A3), and examine applicability beyond adaptive meshes through the cross-mesh evaluation (A4). These controls and clarifications specify the contribution and tested scope of the proposed interaction design.

References

[1] Li et al., Neural operator: Graph kernel network for partial differential equations, arXiv preprint arXiv:2003.03485, 2020.

[2] Wu et al., Transolver: A fast transformer solver for PDEs on general geometries, arXiv preprint arXiv:2402.02366, 2024.

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

This paper proposes GA-AMNO, an adaptive neural operator that separates where to allocate nodes from how their features interact. Physics-derived indicators condition node relocation, while relative coordinates, mesh Jacobians, and state descriptors parameterize diagonal-plus-low-rank edge transformations before attention-weighted aggregation. Spectral and differential residual corrections complete the predictor. The paper introduces operator readability, develops conditional results for consistent aggregation and deformation stability, and evaluates five PDE benchmarks. Validation includes prediction comparisons, component ablations, importance-map and edge interventions, mismatch-stratified feature diagnostics, mesh perturbations, cross-resolution evaluation, and autoregressive rollouts. The full model reports the lowest mean prediction error across the five benchmarks.

Critical strengths
  1. The full architecture reports the lowest mean errors across five PDE benchmarks, with substantial gains on several temporal tasks, supporting its practical utility (Table 1).

  2. Separating allocation and interaction into explicit computational pathways enables targeted interventions, providing a useful framework for investigating adaptive operators beyond aggregate prediction accuracy.

Critical weaknesses
  1. The distinctive contribution and empirical identification of gauge-inspired alignment remain insufficiently established.

    Equations (7)-(10) implement an attention-weighted, geometry/state-conditioned matrix kernel with a diagonal-plus-low-rank parameterization. Matrix-valued, input-conditioned kernels already appear in GKN (Li et al., 2020, Eq. (10)), and MM-PDE (Hu et al., 2024, Appendix H.2) includes state- and geometry-conditioned messages on moving meshes. These precedents leave room for useful specific contributions, but the current comparisons do not establish what additional property the gauge formulation supplies. Theorem 1 addresses a restricted shared-matrix aggregator, and Appendix E explicitly acknowledges that geometry-conditioned edge maps may avoid its limitation.

    Moreover, Eq. (17) rewards transformed source features for approaching target features, mixing representation mismatch with genuine physical variation. For example, with identical frames and hi=1h_i=1, hj=2h_j=2, choosing Ti←j=1/2T_{i\leftarrow j}=1/2 produces perfect compatibility by changing the source value. This illustrates a limitation of the diagnostic; it does not establish that the trained model exhibits this behavior. Although the authors acknowledge the metric's limitations, the conditional theory and edge interventions do not provide a sufficiently discriminating alternative test. Consequently, the evidence supports useful conditional feature transformations more strongly than the claimed interpretation as content-preserving alignment across discretizations.

  2. Allocation effects and predictive gains are not adequately isolated, and essential implementation details are missing.

    The importance weights control displacement in Eq. (6) and directly modulate final features in Eq. (22). Section 4.3.1 does not state whether the feature-modulation path is held fixed during importance-map interventions. Thus, the experiment does not clearly isolate the benefit of relocating nodes. The construction of initial features at relocated coordinates is also insufficiently defined, while the default stride-two reconstruction uses bilinear upsampling of the coarse feature tensor (Appendix D.3). The physical meaning of the adaptive sampling path therefore remains unclear.

    Performance attribution is further complicated because Classic simultaneously removes physics-informed allocation, transport, spectral correction, and differential correction (Appendix C.2). The identity-transport ablation provides useful evidence, especially on Kolmogorov, but removes general edge-dependent channel transformation capacity as well as any alignment effect (Fig. 4). Appendix C.4 additionally lists numerous alignment, cross-mesh, derivative, and spectral losses without their formulas or selected per-dataset coefficients, or clear specification of which supervision is shared with baselines. These omissions affect both reproducibility and attribution. Inconsistent seed reporting between Table 1 and Appendix C.4 also limits assessment of uncertainty, particularly for the small improvement on Darcy.

Questions for authors
  1. How are hi(0)h_i^{(0)} and the state descriptors obtained at the relocated coordinates xiax_i^{a}? How does the default coarse-grid reconstruction account for displacement? During Fig. 5(a), are the Eq. (22) feature gates frozen? If not, can you report displacement-only and feature-gating-only interventions?

  2. What precise structural choice or training constraint distinguishes the proposed transport from existing geometry/state-conditioned matrix kernels? Can you isolate its benefit using a matched alternative with the same geometry/state inputs, prediction heads, applicable training objectives, and comparable capacity? A fixed/adaptive-grid by baseline/proposed-interaction comparison would clarify whether its advantage specifically depends on adaptive discretization.

  3. Can you demonstrate improved agreement for the same physical content represented under different discretizations, while checking that source information is preserved? Please also explain the large increase in transported feature discrepancy for Darcy's low-mismatch group in Fig. 6 and its relationship to prediction error.

  4. Please provide all loss definitions, selected coefficients, and baseline usage, and reconcile the seed counts and standard deviations in Table 1 with Appendix C.4. Also clarify why KS is described as a 256-point one-dimensional problem in Appendix C.1, whereas Table 6 labels training as 64×6464\times64: what resampling and evaluation protocol was actually used?

5 comments
AuthorsOpenReview ↗

Response to Reviewer oiAv (1/4)

W1(a) & Q2. Distinctive contribution of gauge-inspired alignment.

A1(a). GA-AMNO connects where the operator allocates computation with how it exchanges features on the resulting mesh. Graph NO [1] uses input-conditioned matrix kernels, and MM-PDE [2] constructs geometry- and state-conditioned messages on moving meshes. GA-AMNO’s distinction lies in how it exploits the geometry produced by allocation: relative coordinates, endpoint Jacobians, and state descriptors define a structured source-to-target transport map that is applied before scalar attention weighting (Eqs. (7)–(10)). This separation lets us intervene on the transport map while keeping the mesh and attention fixed (Fig. 5(b)). Theorem 1 addresses a shared-matrix aggregator with unchanged scalar weights. As noted in Appendix E, geometry-conditioned maps can instead adapt to each edge’s source and target frames (Eq. (20)). Theorem 2 gives a sufficient condition for consistent aggregation, while Theorem 3 bounds the effects of deviations in attention and transport. The new paragraph “Relation to conditional operator kernels” in Section 3.4 explains these distinctions in detail.

We compare two learned edge-transport parameterizations while keeping the rest of the solver fixed. Structured transport improves both full-field and local spatial accuracy on all three benchmarks. The control replaces only the matrix-generation heads with an MLP that generates a dense transport map from the same geometry and state inputs (Appendix C.4.3; Appendix F.2, Table 5). As shown below, structured transport reduces full-field relative ℓ2\ell_2 error by 5.4% to 11.8% and high-gradient-region error by 5.3% to 15.1%, with lower Laplacian errors across all three benchmarks. The dense control achieves slightly lower spectral error on Navier–Stokes. The spectral metric compares Fourier energy without directly assessing spatial phase, so closer agreement in spectral energy can coexist with the lower spatial errors achieved by structured transport. These results show that the spatial accuracy gains extend beyond the comparison with identity replacement to a learned alternative that also supports edge-dependent channel transformations.

DatasetMean relative ℓ₂: OursMean relative ℓ₂: MLP kernelHigh-gradient region: OursHigh-gradient region: MLP kernelLaplacian: OursLaplacian: MLP kernelSpectral: OursSpectral: MLP kernel
Navier–Stokes0.0031520.0035730.0061520.0072440.0320080.0376620.0014950.001470
Kolmogorov0.0186120.0196650.0240270.0253710.1149230.1166670.0064030.008073
Darcy0.0268270.0284500.0228850.0249630.4350250.4479530.0213520.021964

We compare learned transport with identity replacement on five Darcy meshes. Transport improves prediction even on the regular grid, with the largest gains on the sheared and twisted meshes. For each of five Darcy meshes, we compare gauge transport with identity replacement in the corresponding checkpoint (Section 4.4, Table 2; Appendix C.4.2). Gauge transport reduces prediction error even on the regular grid, showing that its benefit does not require relocation. The gains are larger on every tested nonuniform mesh, especially the sheared and twisted layouts, and the errors with learned transport are close across all five meshes. This pattern is consistent with our design rationale: adjusting incoming features before aggregation is especially valuable when local sampling geometries differ.

Darcy meshIdentity interventionGauge transportRelative error reduction
Regular0.0364620.02793123.40%
Clustered0.0848560.02848166.44%
Sheared0.1416530.02622981.48%
Twisted0.2339450.02660088.63%
Adaptive0.0570720.02682752.99%
AuthorsOpenReview ↗

Response to Reviewer oiAv (2/4)

W1(b) & Q3. Empirical identification of gauge-inspired alignment.

A1(b). Eq. (17) measures message-to-target compatibility on frozen checkpoints; it is not a training objective. In the scalar example, scaling the source feature by ½ reduces the reported discrepancy. But when the two latent features share the same representation frame, the ideal transport is the identity: the source is already expressed in the target frame (Appendix D.2; Eqs. (19)–(20)). GA-AMNO instead learns each transport map through a shared network conditioned on geometry and state. The separate auxiliary loss encourages compatibility alongside field prediction without requiring each message to match the target feature (Appendix C.4.1).

We perturb the source while holding the target fixed. Across all five benchmarks, Gauge’s message response is closer than target copying to a direct encoding of the same perturbation. We hold the target feature, state descriptor, and geometry fixed and apply the same local physical perturbation to the source region under two sampling geometries. On the deformed source grid, we measure how the first-layer transported message changes before attention weighting and aggregation. To obtain the reference response, we directly encode the same physical source region before and after perturbation under the target sampling geometry. The target-copy control returns the fixed target feature and therefore has zero response. We normalize response amplitude by the reference magnitude and compute relative reference error on the full response vector, capturing differences in both magnitude and direction. Across all five benchmarks, Gauge responds to the source perturbation and produces a response closer to the same-source reference than target copying. This shows that the transported message carries changes from the source and quantifies how closely its response matches the direct-encoding reference under the target sampling geometry.

DatasetResponse amplitude: Copy targetResponse amplitude: GaugeRelative reference error: Copy target ↓Relative reference error: Gauge ↓
Navier–Stokes0.0000.7411.0000.319
Rayleigh–Bénard0.0001.1101.0000.273
KS0.0000.5521.0000.521
Kolmogorov0.0000.7441.0000.636
Darcy0.0001.3021.0000.741

In the low-mismatch group of Darcy, the source and target features are already close before transport. The transport map depends on state as well as geometry, so it can still adjust the incoming feature when the local geometries are similar. This can move the message away from the initially close target feature, consistent with the increase from 0.000052 to 0.031156 in Fig. 6. The message contributes to updating the target representation, so a larger distance to its current feature does not imply a larger prediction error against the true solution.

AuthorsOpenReview ↗

Response to Reviewer oiAv (3/4)

W2(a) & Q1. Isolating allocation effects and clarifying the adaptive computation path.

A2(a). In the default reconstruction path, relocation changes which features reach the decoder through message exchange, while input sampling, neighbor connectivity, and decoding sites remain fixed. In the 2D stride-two setting, we encode and average-pool the input, then read and project the state descriptors at the reference node indices to initialize node features. After relocation, the updated relative coordinates and local Jacobians condition both attention and source-to-target channel transformations (Appendix D.2; Eqs. (7)–(9)). The coarse-grid bridge arranges the resulting features by reference index, bilinearly upsamples them, and applies a channel projection (Appendix D.3).

We separate the two intervention pathways in Fig. 5(a) to clarify how the importance map affects prediction. In the original Fig. 5(a) experiment, the feature gates varied with the importance map. To isolate the effect of each pathway, we replay the uniform, shuffled, and shifted interventions on the same 24 test samples. The displacement-only intervention recomputes coordinates while holding the original gates fixed; the gating-only intervention recomputes gates while holding the original coordinates fixed. The table reports percentage changes in prediction error relative to the unmodified model; positive values indicate higher error.

DatasetInterventionDisplacement onlyFeature gating onlyJoint intervention
Navier–StokesUniform−0.0156%+121.87%+121.87%
Shuffled−0.0170%+143.71%+143.72%
Shifted−0.0104%+221.48%+221.50%
KolmogorovUniform+0.0036%+0.70%+0.70%
Shuffled+0.0022%+0.92%+0.93%
Shifted+0.0032%+1.26%+1.26%
DarcyUniform−0.0035%+5.45%+5.45%
Shuffled−0.0045%+5.63%+5.63%
Shifted−0.0045%+5.51%+5.50%

The two interventions distinguish where allocation acts: gating directly controls feature contributions to reconstruction, while displacement changes the geometry used to transform and weight messages before aggregation. The two paths act at different stages: relocation changes relative positions and local Jacobians, which determine how incoming features are transformed and weighted before aggregation; the final gate directly controls how strongly each updated node feature contributes to reconstruction. Fig. 5(a) therefore shows how strongly prediction responds to reweighting the features passed to reconstruction, while the displacement-only comparison isolates its sensitivity to the tested changes in interaction geometry. Fig. 4 addresses the allocation module as a whole: it fixes node coordinates and switches off importance gating and the physical-allocation loss while retaining gauge transport. Together, these comparisons show how allocation shapes both the exchange of features before aggregation and their contribution to reconstruction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Response

Dear PCs, SACs, ACs, and Reviewers,

Thank you for your time and thoughtful feedback, and for recognizing the contributions of our work. Reviewer Y9vi describes separating allocation from interaction as “a relatively clean framing,” and Reviewer Pue4 highlights the “informative mechanism-level interventions.” Reviewer oiAv likewise recognizes “a useful framework for investigating adaptive operators beyond aggregate prediction accuracy” and reports learning “something interesting or useful” (Presentation: 3; Informativeness: 2; Soundness: 2). Building on this feedback, we have conducted extensive additional experiments and expanded the explanations of the method, theoretical assumptions, and evaluation protocol. We have responded to the comments point by point and incorporated the corresponding results and clarifications into the revised manuscript. The main changes are summarized below.

Theoretical scope (Y9vi, oiAv, Pue4). We clarify when geometry-conditioned transport yields representation-consistent aggregation and how deviations in attention and transport affect that consistency.

Node relocation and feature gating (Y9vi, oiAv). We intervene on relocation and gating separately to distinguish their roles: relocation changes the geometry used for message exchange, while gating controls how updated features contribute to reconstruction.

Matched interaction comparison (Pue4, oiAv). We compare structured transport with a dense kernel at comparable capacity and computational cost. Structured transport reduces full-field and local spatial errors on all three tested benchmarks.

Transport across mesh geometries (Y9vi, Pue4, oiAv). Gauge transport reduces Darcy prediction error relative to identity replacement on all five tested mesh layouts, with larger relative gains on the nonuniform meshes.

Source information and compatibility (oiAv). Across five PDE benchmarks, Gauge transport responds to source perturbations more closely to the same-source reference than target copying does. We also clarify that the increased discrepancy in Darcy’s low-mismatch group reflects latent message-to-target distance, not field prediction error.

Fair comparison and reproducibility (Y9vi, oiAv). All methods receive the same observed inputs, and GA-AMNO retains the lowest Navier–Stokes error under matched output supervision. The revision also documents the loss settings, seed and standard-deviation reporting, and the one-dimensional KS evaluation protocol.

We hope the additional experiments and clarifications address the reviewers’ concerns. As of this writing, we have not received further feedback on our responses. We would be grateful if you could consider our responses, the revised manuscript, and the additional evidence alongside the original reviews. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.