acceptodds
Under review as a conference paper at ICLR 2027

Learning Green's Operators with Separable Modulated Kernels

Abstract

Neural Green's operators learn a reusable kernel that specifies how a unit source at one point contributes to a linear PDE's solution at another. Existing basis expansions combine spatial functions that remain fixed as the source point moves, which can require many terms to resolve local features. Motivated by fundamental solutions, we introduce the Separable Modulation Neural Operator (SMNO). We represent the Green's function as a sum of learned offset kernels, each multiplied by separate spatial weights at the source and evaluation points. The kernels depend only on the offset between these points and are shared across PDE instances. Both sets of weights are predicted from the domain shape and the coefficient field. On uniform grids, we apply the operator with fast Fourier transforms (FFTs) without constructing a dense kernel matrix. On Poisson problems with unseen 2D and 3D geometries, SMNO reduces the error on higher-frequency sources by 71–94% relative to basis-expansion Green's operators, and on Darcy flow with unseen coefficient fields it reduces the same error by 82% relative to FNO. A variant with a learned boundary map reduces NGF's error on the Mechanical Components Benchmark by 41%. On a 341×341 grid, building the operator and applying it to 64 sources takes 69 ms, versus 135 ms for FNO, 2.1 s for Transolver, and 10 s for a dense pairwise kernel.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, Green's functions, operator learning, scientific machine learning, PDEs

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_mfqkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Separable Modulation Neural Operator (SMNO), a neural operator architecture for learning Green's functions of linear elliptic PDEs. Unlike existing neural Green's operators that represent Green's functions using finite basis expansions, SMNO parameterizes the Green's function as a sum of learned offset kernels, each modulated by separate spatial weights at the source and evaluation points. The offset kernels are shared across PDE instances, while the spatial weights are predicted from the domain geometry and coefficient field. This representation preserves linearity with respect to the source term and enables efficient operator learning using FFT-based convolutions on regular grids.

The authors evaluate SMNO on Poisson equations with unseen 2D and 3D geometries, screened Poisson problems, and variable-coefficient Darcy flow. The experiments emphasize generalization to unseen source distributions, including localized and higher-frequency sources, where SMNO generally outperforms the evaluated basis-expansion Green's operators and several conventional neural operators. The paper also introduces a separate boundary-map extension that combines a known particular solution with a learned boundary correction for problems with nonzero boundary conditions, demonstrating improvements over existing methods on the Mechanical Components Benchmark.

Critical strengths
  • The proposed method introduces a structured parameterization of Green's functions that explicitly accounts for the relative positions of source and evaluation points.
  • On benchmark problems, such as Poisson equation over 2D channels and Darcy, the proposed method demonstrates substantial improvements over several existing neural operators, particularly on higher-frequency and localized source distributions that differ from those emphasized during training.
Critical weaknesses
  • While the main contribution of the paper is a new parameterization for learning Green's functions, the experiments on the Mechanical Components Benchmark (MCB), which involves complex 3D CAD geometries and nonzero boundary conditions, use a separate boundary-map extension rather than the proposed separable modulated kernel representation. Specifically, it seems a particular solution obtained using a numerical Poisson solver is used in the setting and a network is only trained to predict the remaining boundary correction. Although this approach achieves favorable results compared with NGF [1], the comparison does not directly demonstrate the effectiveness of the proposed SMNO parameterization on this benchmark.
  • Several important methodological and experimental details are not clearly explained in the main text. For instance, the paper provides limited explanation of how quadrature weights are obtained, the computational requirements of the boundary-map extension, and the assumptions underlying the symbol-based approximation in Section 3.2. Although some details are provided in the appendix, clearer explanations would help readers better understand the proposed method and assess its contributions. For example:
    • L50: The introduction contrasts performance on "training-like sources" and "higher-frequency" sources without immediately clarifying how these source distributions are constructed. Although the distinction is explained later in the experimental section and appendix, a brief clarification in the introduction would help readers understand the motivation for evaluating generalization across source distributions.
    • L70–L80: The proposed parameterization is introduced relatively early, but the key advantage of explicitly incorporating the relative offset between source and evaluation points could be emphasized more clearly. In particular, a more direct comparison with the fixed-basis representation used by existing neural Green's operators would help explain why offset kernels are better suited to representing source-dependent local features.
    • Section 3.2: The discussion transitions from the fundamental solution and its correction terms to a symbol-based representation of Green's operators. However, the motivation behind introducing the separable approximation assumed in Eq. (6) could be explained more explicitly. In particular, it would be helpful to clarify under what conditions such a separable representation is expected to approximate the Green's operator effectively, and how this motivates the proposed formulation that jointly predicts ws(x)w_s(x) and vs(y)v_s(y).

Additionally, presentation inconsistencies and incorrect publication information for relevant prior works, such as NGF and Transolver, should be corrected. Please refer to the questions for authors section for details.

Reference

  1. Neural Green’s Functions, Yoo et al., NeurIPS 2025
Questions for authors
  1. Could the authors clarify how the quadrature weights mjm_j are obtained, particularly for irregular meshes? My understanding is that SMNO uses prescribed quadrature weights rather than predicting per-point mass values as in NGF. Does obtaining these weights require mesh information or additional preprocessing at inference time? If so, how does this affect the comparison with mesh-independent approaches?
  2. In L238–L249, the authors describe learning the boundary operator P^\hat{P} while using a box Poisson solver to obtain a particular solution upu_p. Could the authors elaborate on the computational procedure used to obtain upu_p, particularly whether it requires discretizing or meshing the interior of the original domain? Furthermore, could the authors report the prediction error obtained from the particular solution alone, without the learned boundary correction, as an additional diagnostic? This would help clarify how much of the reported improvement comes from the learned correction.
  3. Could the authors explain whether the proposed separable modulated kernel representation in Eq. (7) can be directly applied to the MCB benchmark? Since the reported MCB results use a separate boundary-map architecture, it remains unclear whether the main proposed parameterization offers similar advantages on complex geometries with nonzero boundary conditions. An evaluation that isolates the effectiveness of SMNO itself on this benchmark would strengthen the paper's central claim.
  4. Could the authors clarify the assumptions underlying the separable approximation of the Green's operator symbol in Section 3.2? In particular, is this intended purely as a motivation for the proposed architecture, or is there a broader class of elliptic operators for which the assumed decomposition is expected to provide an accurate approximation?
  5. Minor presentation and reference issues. I also noticed a few minor issues that could be corrected:
    • The formatting of equation references is inconsistent (e.g., "Equation (6)" in L188 versus "Eq. (6)" in L199).
    • The roles of μc\mu_c and ϱc\varrho_c in Eq. (5) could be explained more explicitly.
    • In L261, "Appendices B, A.1, C, and D" could be reordered as "Appendices A.1, B, C, and D."
    • The bibliography should cite Neural Green's Functions as a NeurIPS 2025 paper rather than 2026, and Transolver as an ICML 2024 paper rather than an arXiv preprint. The corresponding citations in tables should also be updated.
3 comments

We thank the reviewer for the detailed comments on the method and the presentation. W1 to W3 follow the weakness paragraphs and Q1 to Q5 the questions.

W1 / Q3 (Direct SMNO on MCB). Yes. We trained SMNO with the kernel of Eq. (7) directly on the MCB data, with no particular solution. It is ahead of NGF and Transolver in all five categories (also General Response, item 1):

fittinggearnutscrewsmotormean
Transolver0.1800.2810.3200.2210.4070.282
NGF0.1600.2430.2750.1890.3380.241
SMNO, direct0.1330.1590.1660.1470.1970.161

The interior block applies the separable kernel over all mesh vertices with the benchmark's lumped masses, and a second separable block integrates the nonzero boundary data over the boundary nodes.

Q1 (Quadrature weights). The weights mjm_j are the quadrature weights of the discretization: 1/N1/N on uniform grids, lumped P1 masses on our generated meshes, and the benchmark's mass matrices on MCB. They are computed from the mesh, or from volume estimates on a point cloud. On MCB they are the benchmark's own mass matrices, which NGF's evaluation also uses, and on our meshes every baseline receives the same weights. Section 3.4 now says this.

Q2 (Particular solution). The direct SMNO of W1/Q3 is trained without any particular solution; upu_p is used only by the boundary-map extension. There, the box solver embeds the vertex cloud into a 1293129^3 Cartesian grid, extends ff via nearest-neighbour lookup, applies a single discrete sine transform, and trilinearly interpolates upu_p back to the mesh vertices. It requires no volume mesh and takes a few milliseconds. On MCB, upu_p alone (with the prescribed boundary values pasted) has a mean error of 7.7, whereas pairing it with an exact harmonic extension yields 0.0086; the revision reports both numbers alongside the 0.141 achieved by the boundary-map extension.

Q4 / W2, Section 3.2 (Assumptions behind the separable symbol). The separable symbol is a classical property that motivates our architecture rather than a formal theorem. Separated representations of this kind are accurate, with SS growing logarithmically in the precision, when the symbol is smooth in xx and belongs to a classical symbol class (Bao and Symes; Demanet and Ying). For the solution operator this holds for second-order elliptic operators with smooth, strictly positive coefficients on smooth domains, whose Green's functions are positive and non-oscillatory. For piecewise-constant Darcy coefficients, the property holds away from interfaces while the learned weights absorb interface jumps; oscillatory kernels like high-wavenumber Helmholtz lie outside this regime. We added this discussion to Section 3.2.

W2, L50 and L70 to L80. The introduction now defines the source families where the distributions are first contrasted, and explains how offset kernels resolve localized, source-dependent features that require many terms in a fixed global basis.

W3 / Q5 (Presentation and references). We added the definitions of μc\mu_c and ϱc\varrho_c immediately following Eq. (5), standardized equation references, reordered the appendix list, and updated the citations for NGF (NeurIPS 2025) and Transolver (ICML 2024).

We hope the direct SMNO result and these clarifications resolve your concerns, and we are glad to discuss any further questions. Thank you again for the careful review.

Dear authors,

Thank you for your thorough response.

My primary concern in the initial review was the lack of clarity regarding the evaluation setting on the MCB-B benchmark. I appreciate the additional evaluation results using SMNO directly, which I believe demonstrate notable improvements over NGF. I also appreciate the further discussion of the separable symbol, as well as the revisions to the manuscript and references.

As my concerns have been adequately addressed, I have decided to increase my rating.

Thank you for taking the time to read our response and revised manuscript. We're glad that your concerns have been addressed, and we appreciate your decision to raise the score!

Reviewer_E1m1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Separable Modulation Neural Operator (SMNO), which represents a Green's function as a sum of shared offset kernels multiplied by instance-dependent spatial weights at the source and evaluation points. The geometry and coefficient encoder is independent of the source, so the predicted operator remains linear in the source. On uniform grids, the operator can be applied through FFTs without explicitly constructing a dense kernel matrix. The paper evaluates source-distribution transfer on 2D and 3D Poisson problems and 2D screened Poisson problems with unseen geometries, and on variable-source Darcy problems with unseen coefficient fields. Kernel ablations and cost measurements examine the benefits of the construction. A separate extension combining a known particular solution with a learned boundary map is evaluated on the Mechanical Components Benchmark.

Critical strengths
  1. The kernel structure and computational implementation form a useful combination. Building on the source-linear operator frameworks of NGF and NGO, SMNO uses shared offset kernels to represent local responses that translate with the source point, while two-sided spatial weights capture instance variation. The number of expansion terms does not directly bound the discrete operator's rank. The structure also supports FFT evaluation and operator caching, providing a practical trade-off between representation capacity, storage, and application cost.

  2. The source-distribution transfer results are persuasive. The main Poisson comparisons use the same source–solution supervision and sampling budgets. Darcy further tests localized and higher-frequency sources on unseen coefficient fields, with three training seeds reported for the main models. SMNO-64 achieves a mean Darcy Band error of approximately 0.088, compared with 0.490 for FNO, while retaining in-distribution accuracy close to FNO. The transfer gains are therefore accompanied by competitive in-distribution accuracy.

  3. The mechanism analysis is substantial and appropriately qualified. Comparisons with a dense pairwise kernel, distance-feature and one-sided-modulation ablations, and linearization, white-noise-coverage, and spectral-truncation diagnostics examine the structural choices from several perspectives. The authors acknowledge that linearity alone does not guarantee accuracy on unseen sources and do not treat the truncation diagnostics as error lower bounds for every model. These distinctions strengthen the empirical conclusions.

Critical weaknesses
  1. Some Poisson comparisons leave uncertainty about the effects of training and implementation choices. The main basis-expansion, DeepONet, and Transolver++ comparisons largely rely on single training runs. A common optimization protocol matches the data budget but does not establish that each architecture has been adequately trained. For example, DeepONet's test error on 3D Fourier sources is close to one, and Transolver++ also has substantially higher errors than the kernel models. Training-convergence or validation evidence would help assess the strength of these comparisons. The three-seed comparison between SMNO and the dense kernel is valuable, but does not address uncertainty in these other baselines. This primarily affects the interpretation of the magnitude of the relative gains.

  2. The mechanical-component experiment does not isolate the contributions of the solver-provided particular solution and the learned module. The extension obtains a particular solution from a box-domain Poisson solver, whereas the cited NGF and Transolver results do not use the same particular-solution computation. The authors explicitly disclose this difference. Consequently, the approximately 41% overall error reduction supports the effectiveness of the hybrid pipeline, but does not independently establish the advantage of the learned boundary map or SMNO kernel structure. A control sharing the same particular solution would help identify the independent gain from the learned boundary map. The comparability of evaluation-node sampling and treatment of prescribed boundary values with the cited baselines also needs confirmation.

  3. The dependence of computational gains on the workload remains insufficiently quantified. The main timing comparison considers 64 sources for the same instance. The paper also discloses that cached basis expansions or dense kernels apply faster on small grids, that SMNO still uses quadratic summation on irregular meshes, and that the larger-grid random-input tests measure cost only. Corresponding comparisons are still missing for frequently changing instances with few right-hand sides, and for conventional solvers that can reuse factorizations or preconditioners. The current evidence is therefore strongest for uniform-grid, multiple-source workloads, and does not yet establish the extent of practical acceleration in broader settings.

Questions for authors
  1. Could you provide training and in-distribution validation convergence evidence for the main Poisson baselines, particularly to explain the high DeepONet and Transolver++ errors? For the basis-expansion controls supporting the main higher-frequency gains, are additional training seeds or configuration checks based on in-distribution validation available to assess the stability of the improvement magnitude?

  2. On mechanical components, is there an alternative boundary-map control using the same solver-provided particular solution, to further distinguish the contributions of the particular solution and learned map? Please also confirm that evaluation-node sampling and treatment of prescribed boundary values are comparable to those of the cited baselines.

  3. Could you report total time for one, a few, and 64 right-hand sides, distinguishing a new instance from repeated applications after caching? Any comparison with conventional solvers should allow them to reuse factorizations or preconditioners, to clarify the workloads in which SMNO offers a practical advantage.

  4. On screened Poisson, SMNO-64 has lower median errors, but some source families have higher 95th-percentile errors than basis expansions or Transolver. Do these single-seed differences suggest difficult regimes associated with particular screening parameters or geometries, or mainly reflect training variability? If relevant stratified analyses are available, please discuss them to clarify the scope of reliability.

5 comments

We thank the reviewer for the supportive review and the precise questions. Q1 to Q4 follow the questions in order; W1 to W3 mark the matching weakness paragraphs.

Q1 / W1 (Convergence evidence and BE seeds). The new appendix figure plots training and Seen-validation curves for all matched Poisson baselines. Every model trains, and the kernel models and BIN keep decreasing until the end of training. The high 3D DeepONet error comes from its index sensors (grid sensors bring it to 0.211), and Transolver++ improves with the standard configuration of its paper (0.066 in 2D) or twice the schedule (0.099); General Response, item 2, lists the new rows. All remain well behind SMNO-64 (0.0042). The gains over BE-1024 hold at every seed:

Band medianBE-1024, seeds 0 / 1 / 2SMNO-64
2D0.0539 / 0.0508 / 0.05310.0155
3D0.9405 / 0.9408 / 0.94690.0550

Q2 / W2 (MCB control and evaluation protocol). The cleanest control is the new direct SMNO, which uses no particular solution at all: 0.161 vs 0.241 for NGF, ahead in all five categories (General Response, item 1). For the boundary-map extension, upu_p alone has a mean error of 7.7, while upu_p with an exact harmonic extension in place of the learned map reaches 0.0086, so nearly all of the extension's 0.141 error comes from the learned map. Evaluation follows NGF: the benchmark's splits and metric (relative L2 error averaged over the test set), computed on every node of each test mesh, with the prescribed boundary values included as in NGF's predictor.

Q3 / W3 (Cost by workload and solvers with reuse). SMNO is faster than sparse LU and algebraic multigrid on both grids, for every number of sources, with and without caching. The new appendix table (General Response, item 4) reports the build time and the time for k=1,4,16,64k = 1, 4, 16, 64 sources on 85285^2 and 4212421^2 grids, so a new instance costs build plus kk sources and a cached one kk sources alone; the solvers reuse their factorization or hierarchy across sources. Total time for a new instance:

Grid, sourcesSMNO-16SMNO-64Sparse LUAMG
85285^2, 13.1 ms3.1 ms16 ms23 ms
85285^2, 643.8 ms6.8 ms31.5 ms569 ms
4212421^2, 15.2 ms7.9 ms0.94 s0.37 s
4212421^2, 6432 ms119 ms3.4 s12.4 s

Q4 (Screened Poisson 95th percentiles). The main reason is training variability, together with a few geometries that are hard for both models; we found no screening regime where SMNO-64 is weaker. We retrained SMNO-64 and BE-256 with two seeds each and stratified per-shape errors by σ\sigma and by geometry:

95th percentile, two seedsSMNO-64BE-256
Fourier (in distribution)0.072, 0.0540.050, 0.101
Band (held out)0.101, 0.1040.126, 0.173

(i) The in-distribution gap flips with the seed. (ii) On held-out families SMNO-64 has the lighter tail at both seeds, and its median is lower in all six families. (iii) The per-shape median error of SMNO-64 is uncorrelated with σ\sigma (Spearman +0.08 and -0.06), and the same geometry has the largest 95th percentile in all four runs. Transolver's lighter in-distribution tails come with 95th percentiles of 0.92 on Band and 0.44 on Localized.

We hope this answers your questions, and we are glad to discuss anything further. Thank you again for the constructive feedback.

Thank you for the additional experiments in the General Response and the detailed replies. The direct SMNO results on MCB, additional BE seeds, and workload-dependent timings substantially address my main concerns.

Please clarify the CPU/GPU hardware differences and accuracy conditions alongside the main-text solver timing comparison, and resolve the seed-count inconsistency for direct SMNO between Table 2(b) and Appendix F.

Thank you for the follow-up. We are glad the new results address your main concerns. The updated revision resolves both remaining points.

Solver timing comparison. Section 4.3 now states the hardware next to the comparison: the sparse LU and multigrid solvers run on one CPU core, and SMNO runs on one H100 GPU. We use the widely used implementations of the two solvers, SuperLU through SciPy and PyAMG, which are CPU codes. We also timed the solvers with 32 threads (Appendix G): only the 64-source LU solve on 4212421^2 gets faster (2.5 s to 0.9 s), so SMNO remains faster in every setting. The same sentence in Section 4.3 gives the accuracy of each side: the solvers are exact to solver tolerance, and SMNO has the errors of Table 1(b).

Seed count for direct SMNO. We trained two more seeds of the direct SMNO in every category. Its column in Table 2(b) now reports three-seed means, and the caption and Appendix F agree. The mean over the five categories is 0.161, 0.156, and 0.157 at seeds 0, 1, and 2, against 0.241 for NGF, and each of the 15 runs is below NGF in its category:

fittinggearnutscrewsmotormean
NGF0.1600.2430.2750.1890.3380.241
SMNO, direct, three seeds0.1290.1560.1590.1520.1940.158

Our earlier replies quote the seed-0 values (mean 0.161).

Thank you again for the careful reading.

Reviewer_2ucyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose Separable Modulation Neural Operator (SMNO), a novel model that represents Green’s function as a sum of learned offset kernels. We test it on the Poisson problem on a regular grid, comparing the proposed method with baselines including DeepONet. Transolver and FNOs. Moreover, it is tested on Poisson problems with unseen 2D and 3D geometries, still compared with other baselines like Transolver and NGF.

Critical strengths

The parametrization proposed by the authors makes a clear contribution to achieving better generalization outside the training distribution; the theoretical presentation gives a reasonable intuition, and it is clearly empirically proven.

The method is evaluated on several scenarios, such as 2d and 3d Poisson problems with held-out geometries, varying coefficients, and varying sources.

Critical weaknesses

Some baselines seem to be not well trained. For example, DeepONet in 3D has a relative L^2 of ~1, which should be close to an untrained network. Another example is Transolver++, which seems to be underperforming with respect to the original paper. Clearly cannot be directly compared to the reported numbers in the original paper but, on a similar problem (Darcy), achieves almost two orders of magnitude lower relative L^2. Moreover, if I understood correctly, BE-64, BE-256, ... are not the original model, but the authors' adaptation, so claiming: "SMNO reduces the error on higher-frequency sources by 71–94% relative to basis-expansion Green’s operators" to me seems not to be empirically proven.

In Table 2 (b), the comparison seems to be unfair; Transolver and NGF learn the whole mapping, whereas the boundary map needs to solve an easier task.

If I am understanding correctly: On Poisson, DPK has lower mean errors than SMNO, and it also has lower errors in Figure 3(a,b). Moreover, DPK is omitted from Table 1(a). SMNO’s advantage over DPK appears only on screened Poisson and Darcy. This does not match the abstract and introduction, where shared offset kernels with two-sided modulation should be the main advantage of the model.

On a regular grid, there is a missing baseline [1], which it is known to approximate well elliptic problem.

The representation is a known structure, and given the existence of NGF and the narrow scope of the method (the method works only on linear elliptic problems with Dirichlet conditions and non-oscillatory Green’s functions), overall it seems an incremental improvement.

[1] Sittoni et al., Neural-HSS: Hierarchical Semi-Separable Neural PDE Solver, ICML 26.

Questions for authors

See weaknesses.

3 comments

We thank the reviewer for the careful reading and the specific questions on baselines, which led to most of the new controls in the revision. We number the weakness paragraphs W1 to W5 in order.

W1, part 1 (DeepONet in 3D, Transolver++, Darcy). Both baselines are well trained; their gap to SMNO is structural (General Response, item 2, with the new training curves). 3D DeepONet's error of 1.00 comes from its standard branch, which reads the source at fixed array indices. On an unseen shape these indices fall at different places, so the best prediction for a zero-mean source is zero. With grid sensors it reaches 0.211. For the attention models we added Transolver++ in its paper's standard configuration and the released Transolver. The best of them, the released Transolver, still has 3.4 to 9.5 times the error of SMNO-64 in distribution and 12 to 26 times on Band. On Darcy, the published 0.0057 is for the single source f=1f = 1, while our task varies the source and tests held-out source families. On our in-distribution sources Transolver reaches 0.019 and Transolver++ 0.022, and our solver reproduces the public f=1f = 1 labels to within 0.013%.

W1, part 2 (BE controls and the 71 to 94% claim). NGF released code only for its 3D benchmark, with no 2D version, so we built the BE controls with NGF's structure: the kernel is a sum of RR products pi(x)qi(y)p_i(x) q_i(y) of learned point features, with the same inputs and data as SMNO. The one change is separate heads for p and q, which only enlarges the family; on Darcy, NGF's shared-feature form gives the same Band error (0.865 and 0.855 at R=128R = 128 and 256256, vs 0.845 for BE-512). The abstract now says "relative to NGF-style basis-expansion controls", and against NGF itself, on its own benchmark, SMNO reaches 0.161 vs 0.241 (General Response, item 1).

W2 (Table 2(b) and the boundary map). We trained SMNO directly on the MCB (f,g,u)(f, g, u) triples with no solver-provided particular solution (General Response, item 1). It reaches 0.161 vs 0.241 for NGF and 0.282 for Transolver, outperforming NGF in all five shape categories. The boundary-map version is retained and explicitly labelled as an extension.

W3 (DPK in Table 1(a)). We added DPK to Table 1(a) (General Response, item 3). DPK is our own control: the same model with a dense pairwise head in place of the separable kernel. Similar accuracy on constant-coefficient Poisson is therefore expected, and there the two are within seed variability. SMNO is far cheaper, and much more accurate where the PDE instance changes the kernel:

DPKSMNO-64
Build at 3412341^210 s, 54 GB2.8 ms
Build at 4212421^2out of memory4.5 ms
Screened Poisson, Band error0.1070.018
Darcy, Band error0.2050.088

Sections 4.2 and 4.4 now state this.

W4 (Neural-HSS baseline). We implemented Neural-HSS from the paper and trained it on variable-source Darcy (General Response, item 5):

DarcySeenLocalizedBand
Neural-HSS0.01340.09471.855
SMNO-640.01340.01190.0877

Matching SMNO-64 on Seen shows that the implementation trains well. It falls behind on held-out families, consistent with its design for a single operator on a fixed grid, where the coefficient enters only as an input channel. The appendix states this, and the related work section now contrasts the architectures.

W5 (Scope and contribution). While classical separable symbols motivate our approach, prior learned Green's operators have not combined offset kernels with two-sided modulation. Within the elliptic Dirichlet class targeted by NGF, NGO, and basis expansions, this reduces held-out source error by an order of magnitude (3D Band 0.055 vs 0.94 for BE-1024; Darcy Band 0.088 vs 0.85 for BE-512). The same structure keeps the operator cheap: on a 3412341^2 grid, building it and applying it to 64 sources takes 69 ms, vs 135 ms for FNO, 2.1 s for Transolver, and 10 s for a dense pairwise kernel, and it stores O(SN)O(SN) values where the dense kernel needs 54 GB. The conclusion clarifies the scope: screened Poisson already varies the differential operator, Neumann and Robin conditions only alter the boundary block, and the architecture applies to any operator admitting a parametrix of the form in Eq. (5).

We hope the revision resolves your concerns, and we would be glad to discuss any further questions. Thank you again for your time and the constructive review.

Reviewer_2ucyOpenReview ↗

Score update

I thank the authors for their response and the additional results. The authors addressed my concerns. Therefore, I will raise my score by 1 point.

Thank you for reading our response and the revision, and for raising your score! We are glad the additional results addressed your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their detailed feedback. The revision (changes in red) adds the following experiments and controls; the individual replies refer to these items by number.

  1. Direct SMNO on the Mechanical Components Benchmark (all reviewers). We trained SMNO directly on the benchmark's source, boundary, and solution data instead of using the boundary map, with 16 offset kernels and the benchmark's splits and metric. On the unseen-shape, unseen-problem split it beats NGF and Transolver in every category:
fittinggearnutscrewsmotormean
Transolver0.1800.2810.3200.2210.4070.282
NGF0.1600.2430.2750.1890.3380.241
SMNO, direct (new)0.1330.1590.1660.1470.1970.161

The boundary-map version stays in the paper, labelled as an extension.

  1. Poisson baselines (2ucy, E1m1). A new appendix figure gives training and validation curves for every matched Poisson baseline; all of them train (3D DeepONet's training loss falls from 0.88 to 0.16). Table 1(a) keeps the submission rows and adds three in red. The submission ran Transolver++ from its released code in the only configuration the release ships, the airplane preset; the new row uses the configuration its paper gives for standard benchmarks. The released Transolver, our baseline on screened Poisson and Darcy in the submission, now also appears in Table 1(a) next to its successor. The submission's DeepONet is the standard form, which reads the source at fixed indices of the nodal array; on a new shape these indices fall at different places, which is why its 3D error is near 1. The new row reads the source on a fixed grid over the bounding box (Fourier/Band):
Row2D3D
DeepONet (submission)0.100/1.0491.001/1.012
DeepONet, grid sensors (new)0.093/1.0530.211/1.009
Transolver++, airplane preset (submission)0.346/1.3860.344/1.484
Transolver++, standard config (new)0.066/0.5420.328/0.991
Transolver, released code (new)0.040/0.4040.068/0.658
SMNO-640.0042/0.01550.0198/0.0550

The best attention model, the released Transolver, still has 3.4 to 9.5 times the error of SMNO-64 in distribution and 12 to 26 times on Band. The reason is structural: our benchmark tests unseen geometries on new sources, so a model has to learn the Green's operator of each geometry. Kernel models are linear in the source by construction, so one learned kernel serves every source. Transolver and DeepONet pass the source through a nonlinear map and must learn superposition from data; none of them gets below 0.40 on Band. The NGF paper finds the same for Transolver on 2D Poisson (Yoo et al., Table 1: 0.053 on training functions vs 0.372 on unseen ones). On MCB, by contrast, each shape has only 16 training problems and every method pays for extrapolating from them, so the methods stay within a factor of two.

  1. DPK in Table 1(a) (2ucy). DPK is the dense version of our own model, with a pairwise MLP head in place of the separable kernel. On constant-coefficient Poisson it is slightly more accurate than SMNO, within seed variability, as expected from a dense kernel. It costs far more: at 3412341^2 it takes 10 s and 54 GB to build, vs 2.8 ms and O(SN)O(SN) storage for SMNO-64, and at 4212421^2 it does not fit on the GPU. On screened Poisson and Darcy, where the instance changes the kernel, SMNO is more accurate (0.018 vs 0.107 and 0.088 vs 0.205). Sections 4.2 and 4.4 state this.

  2. Cost by workload (E1m1). A new appendix table reports build and apply times for 1 to 64 sources on 85285^2 and 4212421^2 grids, against sparse LU and algebraic multigrid solvers that reuse their factorization or hierarchy.

  3. Neural-HSS (2ucy). We implemented Neural-HSS from the paper and trained it on variable-source Darcy under the Table 1(b) protocol. It reaches Seen 0.0134, Localized 0.0947, and Band 1.855 (SMNO-64: 0.0134, 0.0119, 0.0877). The row is in Table 1(b) and the details in the appendix.

  4. Released code for Transolver rows. All Transolver and Transolver++ rows use the authors' released code except Darcy Transolver++: the Transolver++ release contains only the million-point airplane pipeline, so that row implements the Transolver++ equations inside the released Transolver Darcy architecture. The released Transolver++ file, trained on Darcy under the same protocol, is weaker on Seen and Band (appendix), so Table 1(b) keeps our implementation.

  5. Further additions. BE-1024 at three seeds and a per-shape analysis of the screened Poisson tails (E1m1); clarifications on quadrature weights, the particular solution, the scope of the separable symbol, μc\mu_c and ϱc\varrho_c, and the source families (mfqk); reference updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.