acceptodds
Under review as a conference paper at ICLR 2027

Structure-Factor Coordinates for Finite-Aperture Operator Learning

Abstract

Learning wave responses from aperture geometry requires resolving interference between spatially distributed elements. We introduce a raster-free physical interface that represents this interference through point-set Fourier intensity and multiscale angular moments of the structure factor. In paired acoustic controls, the interface reduces response MAE by 22.7% for Transolver-MS and 21.4% for Point-DeepONet. When the square-triangle family is entirely withheld, the gains persist across three seeds: 10.2-18.2% for Transolver-MS and 18.2-20.5% for Point-DeepONet across two training settings. SURFNO complements the same interface with a fitted kernel transform and learned residual, and performs competitively with spectral baselines. Antenna and scalar diffraction benchmarks reuse the Fourier-intensity construction under different observation measures, while solver-verified inverse design separates representation quality from feasible mask-space search. Equal-spectrum counterexamples delineate the information discarded by spectral compression. Together, the results establish structure-factor coordinates as a reusable physical interface across predictor architectures and an unseen aperture family; finite-window observations and geometry-dependent coupling require complementary geometric or source-state information.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
structure factors, operator learning, physics-informed representation learning, geometric generalization, wave apertures, inverse design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SwZ7OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes structure-factor coordinates, an input representation for predicting the frequency response of finite-aperture wave systems from the active element coordinates. The structure factor S(q)S(q) is computed directly from the point set and compressed into radial spectra and angular moments at multiple scales. It is then appended as an additional input to point-based predictors (Transolver-MS, Point-DeepONet), and to the proposed model named SURFNO that combines a fitted kernel transform with a learned residual.

Critical strengths
  • The structure factor connects the element layout to the wave response through a known physical relation, which leads to improvements in two different predictors.
  • S(q)S(q) is computed directly from the point set, which avoids the sensitivity to rasterization observed for the CNN baseline.
Critical weaknesses
  • Readability. The paper needs substantial revision to improve readability.

    • The paper contains many unnecessary and defensive phrases throughout the paper, which do not convey meaningful information but hinder understanding. A few examples: "MS means multiscale, not multiple seeds", "This is reuse of a physical representation, not evidence that ...".
    • Claims are often raised and then immediately walked back, making the main message difficult to identify.
    • Essential details are missing, including the Transolver-MS and Point-DeepONet architectures, S(q) fusion and readout, original-model preliminaries, and the detilas of S(q)S(q) (qq range, bins, harmonic orders).
  • Unclear contributions. It is difficult to identify the core contribution of the paper.

    • The proposed "physical input interface" appears more like an additional input feature: it is the classical structure factor, which the authors acknowledge is not novel, compressed by standard radial/angular averaging and appended as an input feature.
    • The other two listed contributions provide limited empirical support for this feature and mostly negative findings.
    • SURFNO, the only new model, is neither listed as a contribution nor the best-performing predictor.
  • What is reusable, and why? The paper does not establish what makes the representation "reusable" or what benefit this provides.

    • S(q) is a deterministic function of the coordinates already given to the models, serving as an additional hand-crafted global descriptor.
    • The representation itself appears not reusable in practice: each model is modified to accept S(q) and trained separately, while the cross-physics experiments also fit separate models.
    • Experiments do not clearly demonstrate the advantages of sharing the representation.
  • Insufficient evaluations. The experiments do not sufficiently support the main claims.

    • The experiments mainly test whether S(q)S(q) improves prediction as an additional input, not whether it is reusable.
    • Even as a feature ablation, Tab. 2 only removes S(q)S(q) from the authors' modified models. Comparisons with the original models adapted with a minimal output modification is missing, and the geometry-only Transolver is reported in Tab. 9 but its accuracy is not compared against the proposed models.
    • In inverse design, the search procedure appears more important than the representation: SURFNO and CNN perform on par under the same search, while the path directly using S(q)S(q) performs worst. These results contradict with the paper's main claims and need more thorough analysis and explanations.
Questions for authors
  • What exactly is reused beyond the formula for computing S(q)S(q)? Can the authors provide an experiment demonstrating a benefit of reuse, e.g., transferring a trained S(q)S(q) encoder across predictors or physical systems?
  • How do the original Transolver and DeepONet, adapted with a minimal output modification and using coordinates alone, perform on the acoustic benchmark? What is the accuracy of the geometry-only Transolver in Tab. 9?
  • In inverse design, SURFNO and CNN perform on par under the same search, while the path directly using S(q)S(q) performs worst. Could the authors provide deeper analysis on these results and connections of these results to the advantages of the reused representation?
3 comments

Thank you for the four concerns. We revised the presentation and added input, original-model and encoder-transfer controls. References use revised numbering.

1. Readability and missing details. Section 3.1 now explains points → intensity → radial/angular channels → multiscale tensor → predictor. It defines reciprocal ranges, seven scales, 64 radii, 32 azimuths, orders 2/4/6/12 and the 7×9×64 tensor. Appendix B.2/Table B2 specify geometry/spectral encoders, fusion and readout dimensions. Tables B1/B3 collect physical/data parameters and training protocols. Repeated defensive phrasing is removed; Appendices A–E organize derivation, implementation, prediction, reuse and design.

2. Contribution and SURFNO. Structure-factor coordinates compress geometry-derived relative-phase information into a common multiscale directional input for different predictors.

The added three-seed SF/FFT comparison holds predictor capacity, geometric inputs and the 120-epoch budget fixed, using a common fixed IDCT adapter:

PredictorSF MAE (dB)FFT MAE (dB)Reduction
Transolver-MS0.11196 ± 0.001460.16113 ± 0.0018430.5%
Point-DeepONet0.11928 ± 0.002520.15859 ± 0.0016224.8%

Means ± sample SD; every pair improves MAE/RMSE (Table 2). The comparison includes sampling/preprocessing and irregular-template raster collisions. A separate from-scratch unseen-template test improves both predictors while retaining a generalization gap (Table 4).

SURFNO is an interpretable baseline: its fitted response kernel and learned residual are inspectable separately. Its purpose is physical interpretation, not necessarily best accuracy. Coupled-acoustic coefficients are effective mappings rather than uniquely identified mechanisms. Fixed-budget with/without-core MAEs of 0.0986/0.0974 dB do not show a core accuracy gain (Appendix C.4).

3. What is reused? We transfer 159,680 spectral-module parameters from seed-0 Transolver-MS to Point-DeepONet; geometry processing and readout remain target-specific. Four conditions pair other initializations across three target seeds, using the same acoustic split, 120 target epochs and validation-RMSE selection:

Target spectral modulesMAE (dB)RMSE (dB)
Random/trainable0.113531 ± 0.0222220.265181 ± 0.008294
Transferred/fine-tuned0.116094 ± 0.0042920.260902 ± 0.003723
Random/frozen0.152818 ± 0.0034200.364129 ± 0.006904
Transferred/frozen0.098629 ± 0.0003900.244025 ± 0.000864

Frozen transfer improves every target seed: 13.1% lower mean MAE than random/trainable and 35.5% lower than random/frozen. Fine-tuning does not improve mean MAE (Table 3; Appendix D.1). Target training falls from 1163 to 359 s; including 1303 s source training gives 1662 s for one source-plus-target fit. This is same-task cross-predictor reuse with one source encoder, not cross-physics transfer. Antenna/diffraction models are fitted separately.

4. Original models and geometry-only accuracy. Submitted Table 9's geometry-only Transolver achieves 0.13818 dB MAE, with 365,933 parameters and coordinates, layout and 11 descriptors. Revised Table C5 now reports its accuracy. New coordinates-only controls use three seeds, 120 epochs and validation-RMSE selection:

ControlParametersMAE (dB)
Original Transolver + global readout211,4230.190144 ± 0.005139
Classical MLP DeepONet + coordinate packing780,4170.809484 ± 0.110168

Table C3 specifies adaptations and per-seed results. DeepONet selects epochs 118/119/120; its result describes this configuration, not a convergence limit. These architecture controls differ in capacity. Direct input evidence comes from matched-capacity SF/FFT and full/zero studies. Continued full/zero now covers three seeds, each arm continuing its own independently initialized lineage with historical early stopping (Table C2).

5. Inverse design. Across 72 antenna targets, spectrum-mediated paths achieve median exact-evaluator MAEs of 3.44×10⁻⁵/3.41×10⁻⁵ versus 2.39×10⁻⁴ for the strongest direct-logit path; differing budgets make this a complete-procedure comparison. Across 30 acoustic targets, matched CNN/SURFNO swap search gives 0.0750/0.0771 dB, versus 0.2031 for spectrum-first and 0.2730 for continuous-mask search (Figure 3; Table 5).

Coupled amplitudes, finite windows and discrete realization add dependencies beyond Fourier intensity. Antenna results support spectral design; acoustic results favor feasible geometry refinement. Forward gains do not imply inverse gains for every predictor (Appendix E).

Thank you for the detailed response. Unfortunately, my concern regarding the contribution and reusability still remains.

The revised version now lists contributions as 1) S(q)S(q), 2) encoder transferability on the same task, 3) SURFNO as an interpretable baseline. In my view, using S(q)S(q) as input offers little new insight and seems to have limited novelty. Second, it would be more convincing if the authors showed what advantages and findings "transferring the encoder on the same task" could give. Finally, SURFNO doesn't give improvement in terms of accuracy, while the interpretability claim needs further discussion and validation.

Regarding the reusability, it's difficult to claim that a representation directly derived from a well-known physical formulation is a "reusable representation", while the encoder transferability on the same task needs more discussion, particularly in terms of its advantages.

I appreciate the efforts to improve readability and provide additional results. I'm raising the score, but I still lean toward rejection for the reasons above.

AuthorsOpenReview ↗

Thank you for reconsidering the paper

Your comments help us clarify how the representation, encoder-transfer experiment and interpretable baseline support one central contribution.

Contribution. We construct multiscale, directional coordinates that organize global interference information into a common input for different geometry-to-response predictors. This allows us to investigate whether the spectral features learned within one architecture remain useful when its geometry encoder and readout change.

Advantages and findings of reuse. On the same acoustic task, transferring one frozen Transolver-MS spectral encoder to Point-DeepONet reduces mean MAE by 13.1% relative to training from scratch and by 35.5% relative to a randomly initialized frozen encoder, across three target seeds (Table 3). The latter control establishes that the learned weights carry useful information. Freezing also gives the lowest mean MAE among the four initialization/training conditions, while fine-tuning does not improve mean MAE over ordinary training. Thus, the experiment demonstrates that useful spectral features can survive an architectural change without further adaptation. With the source encoder available, target training takes 359 seconds versus 1,163 seconds from scratch. Source training costs 1,303 seconds.

Physical interpretability. We agree that the acoustic core ablation shows no accuracy gain from adding the core. Its interpretation can be assessed more directly in Figure A1, which compares physical and fitted kernels. The fitted spectral bands shift toward higher radial wavenumbers as antenna frequency increases, and toward lower wavenumbers as diffraction wavelength increases, following the respective physical relations. These observations connect the fitted weights to physically meaningful spatial scales. SURFNO makes this spectral contribution and the learned correction separately inspectable. For coupled acoustics, the fitted core represents an effective spectral-to-response mapping.

Together, these results show how the proposed coordinates support learned-feature reuse across two architectures and permit physical inspection of the spectral-to-response relation.

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

The paper considers modelling wave responses from aperture geometry using a neural operator. In this problem, the target values are also a function of the pairwise interference between the aperture elements. While this interactions can be learned, this paper proposes to provide them to the input of a neural operator instead, thus injecting inductive bias. The paper proposes a way to construct and featurize this information using the normalized directional structure factor. It then shows that this improved featurization helps improve model performance across different neural architectures.

Critical strengths

The paper identifies a key component in wave systems, and propose a new way to incorporate pairwise interference information into current models that is applicable and effective with different architectures.

Critical weaknesses

Presentation. From what I can tell, the main contribution of the paper is described in Section 3.1 which is very difficult to parse: The method is not well explained, nor well motivated. In addition, there are notational overloads (line 097 re-defines S_m) and missing notation definitions (q is not defined, just used). This poor communication of the key contribution appears to be a major weakness of the paper.

Questions for authors

Have previous papers on wave systems tried to incorporate information using the structure-factor?

3 comments

Thank you for recognizing the physical motivation and applicability across architectures. We have addressed presentation through a rewritten Section 3.1, corrected notation, expanded Related Work and clearer experimental organization. References below use revised manuscript numbering.

1. Physical idea and presentation. Relative element displacements determine phase differences and hence constructive or destructive interference. Structure-factor coordinates summarize this geometry-derived information across scales and directions in a common input tensor for different predictors.

Section 3.1 now follows point coordinates → directional Fourier intensity → radial/angular channels → multiscale tensor → predictor, with definitions before use and a two-element example. Figure 1 illustrates the pipeline. Section 3.2 separately introduces SURFNO as an interpretable fitted kernel-transform core plus a learned residual. Appendices A–E group physical formulation, implementation, prediction controls, reuse/generalization and inverse design.

2. Notation and implementation. The directional function and its radial mean have separate symbols:

\frac{\left|\sum_{i=1}^{N_m}w_{m,i}e^{-\mathrm{i}\mathbf q\cdot\mathbf r_{m,i}}\right|^2} {\left(\sum_{i=1}^{N_m}|w_{m,i}|\right)^2+\epsilon}, \qquad \overline S_m(q)=\frac{1}{A}\sum_{a=1}^{A}S_m^{\mathrm{dir}}(q\mathbf e_a).$$ Here $m$ indexes the sample, $\mathbf r_{m,i}$ is an active position, $w_{m,i}$ its prescribed weight, $\mathbf q$ a reciprocal vector, $q=\|\mathbf q\|$, and $\mathbf e_a$ an angular sampling direction. Positions use metres and reciprocal coordinates inverse metres; $\mathbf q\cdot\mathbf r_{m,i}$ is a phase. Response frequency is a separate query. Nonempty unit-weight acoustic inputs use $\epsilon=0$ and normalization $N_m^2$; logarithms and angular normalization use a separate floor $\epsilon_s=10^{-12}$. The sampling paragraph states seven scales $\alpha\in\{1,2,4,8,16,32,64\}$, 64 radii per scale from $10^{-6}$ to $\alpha\pi/D$, 32 azimuths and harmonic orders 2, 4, 6 and 12. Here $D$ is the smaller full-template span. Radial and real/imaginary moment channels form a $7\times9\times64$ tensor. The predictor-integration paragraph explains encoding, fusion and frequency-query readout; Appendix B.2/Table B2 give layer dimensions. **3. Previous learning studies using structure factors.** Yes, related studies exist. Expanded **Section 2** now distinguishes their tasks and representations: - [Chang et al. (2022)](https://www.nature.com/articles/s42005-021-00778-y) infer effective interaction potentials in colloidal suspensions from scattering spectra. - [Munshi et al. (2022)](https://www.nature.com/articles/s41524-022-00939-9) reconstruct quantitative structure-factor images from electron diffraction for strain mapping. - The [Arnau et al. research poster](https://www.upv.es/contenidos/ENCDOC/info/U0957508.pdf) learns mappings between one-dimensional point configurations and structure factors for hyperuniform design. Our target is a finite aperture's wave response. We directly compute normalized directional intensity from its point set, organize multiscale radial spectra and angular moments, and supply the tensor to different response predictors. To our knowledge, the studies above do not present this particular coordinate construction for finite-aperture geometry-to-response learning. “Coordinates” denotes a feature-space description indexed by reciprocal scale, radial sample and angular order. The novelty concerns this learning representation and its demonstrated uses, while the structure-factor definition is classical. New three-seed SF/FFT comparisons lower mean MAE by **30.5% in Transolver-MS and 24.8% in Point-DeepONet**, with lower MAE/RMSE in every pair (Table 2). Transferring and freezing a Transolver-MS spectral encoder in Point-DeepONet lowers mean MAE by **13.1%** over random trainable initialization (Table 3). This demonstrates same-task cross-predictor reuse; fine-tuning does not improve average MAE. The revised text separates this evidence from sharing only the feature formula.

Thank you for updating the paper. I think section 3.1. is improved, however, it reads like a feature engineering recipe - it states what is being done, without motivating why, which limits the scientific insight that can be extracted. In my assessment, the paper needs to be written in a more accessible way for the audience of ICLR, therefore, I will keep my score at 2. At the same time, as indicated, my confidence is 1.

AuthorsOpenReview ↗

Thank you for rereading the revision and for clarifying your concern

We agree that the design rationale needs to be explicit for readers across the ICLR community. We have rewritten Section 3.1 around the learning problem, the physical basis for the representation, and the evidence supporting its use.

Why provide these coordinates? Predicting a wave response requires accounting for interference between spatially separated elements. Squaring the summed wave amplitude produces a cross term for every element pair. The structure factor collects these terms directly from their relative positions. Our learning hypothesis is that organizing this global information explicitly gives different predictors a useful inductive bias. Section 3.1 now begins with this hypothesis and derives the connection to pairwise interactions before introducing the compressed features.

Why radial, directional and multiscale features? The observation law guides what information to retain. Radial intensity determines an azimuthally averaged response under the stated propagation and source assumptions. Directional observations also depend on orientation. A short two-element example shows how rotating a layout preserves its radial mean but changes its directional response. Angular moments retain this variation. Element separations also determine the spacing of spectral oscillations, motivating overlapping reciprocal ranges that balance sampling resolution and coverage. The particular ranges and harmonic orders are empirical choices. The revised text explains these choices alongside the assumptions of the underlying physical relation.

How is this reasoning tested? The seed-0, capacity-matched channel ablation reduces MAE from 0.1407 to 0.1044 dB when angular information is retained (Fig. 2b). The three-seed comparison with the tested raster-FFT pipeline reduces mean MAE by 30.5% in Transolver-MS and 24.8% in Point-DeepONet (Table 2). These results support the usefulness of the complete input construction across two architectures. Transferring and freezing one trained Transolver-MS spectral encoder in Point-DeepONet also improves mean MAE by 13.1% over random trainable initialization across three target seeds on the same acoustic task (Table 3).

The revised Sections 3.2–3.3 explain how the observation relation motivates SURFNO's interpretable fitted core and why finite-window, coupled-acoustic prediction also uses geometry. The equal-spectrum examples in Section 4.5 demonstrate how different layouts can have identical directional intensity yet different finite-window responses. Together, these explanations connect the representation to both its predictive benefits and its physical scope.

We would appreciate your assessment of whether the revised Section 3.1 addresses your concern about scientific motivation and accessibility.

Reviewer_HpQ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper give cheat sheet to AI along with patterns (wave). The cheat sheet summarizes the spacing between every pair of speakers, and in which direction. Physicists call it the structure factor. They have used it for about a century to study crystals with X-rays. The tool is old. The new part is using it as an input for AI. Although I am not expert in this particular topic, I have focused more on the DL component. The paper made 30,000 random patterns on there grid types, they used AI with and without cheat sheet. Cheat sheet allowed AI to reduce errors, but is not that expected? They also did surprise test, by hiding one whole grid type during training, and test full of mixed ones, that is similar to denoising AE. When they add physical formula, then results improved further.

Critical strengths

-- experiments seem to be designed carefully although I am not expert in this particular domain of physical space and waves. -- paper is honest about negative results too -- interesting problem to solve

Critical weaknesses

-- Overall, the paper shows that adding a classical physics feature improves two existing networks. It does not show anything beyond that, and that finding alone does not meet the ICLR bar for novelty or significance. -- angular basis is capturing symmetry maybe? this has not been shown. -- novelty seems to ve quite limited. Giving cheat sheet to AI or helping with prior, and doing surprise test (denoising AE) are all existing methodologies. Application maybe new. -- zero-input control does not test representation specificity. -- simulator and dataset are underspecified.

Questions for authors

Please rerun the same Transolver-MS and Point-DeepONet pairs with a matched-dimension alternative input: a pair-distance histogram, raster FFT magnitudes of the 50×50 mask, or the analytical uncoupled response computed from coordinates with equal weights. Does the structure-factor tensor outperform these alternatives? If it does not, what is specific to your interface?

What is the zero-spectrum MAE at the continued stage, across all seeds? Was the Point-DeepONet zero-spectrum arm trained from scratch or continued from a checkpoint trained with spectra? Appendix B argues that a continuation intervention measures dependence, not benefit.

Please state the medium, sound speed and element size.

3 comments

Thank you for the careful assessment. We have clarified the contribution and added controlled experiments. References below use revised manuscript numbering.

1. Contribution and task. Structure-factor coordinates expose geometry-derived relative-phase information through a fixed-size, multiscale radial-and-directional tensor. The contribution is this input construction and its demonstrated use across predictors; the underlying structure factor is classical. Section 2 distinguishes our geometry-to-response task from scattering inversion, structure-factor reconstruction and hyperuniform point-set design. Evidence now includes an FFT alternative, unseen-template prediction and learned-encoder transfer; frozen Transolver-MS-to-Point-DeepONet transfer improves target MAE by 13.1% on the same acoustic task (Table 3).

The holdout trains predictors from scratch on the square layout template alone or both square and triangle templates, withholding the entire square–triangle family (Table 4). It predicts responses from unseen geometries; it is not denoising-autoencoder reconstruction.

SURFNO provides an inspectable fitted kernel-transform core plus a learned residual (Section 3.2). The fixed-120-epoch with/without-core MAEs are 0.0986/0.0974 dB, so that experiment does not show an accuracy gain from the core. Removing a previously used core during continuation is a separate intervention (Appendix C.4).

2. Angular information. Translations cancel in Fourier intensity; rotations rotate its directional spectrum. Adding angular channels lowers capacity-matched Transolver MAE from 0.1407 to 0.1044 dB. Seven scales give 0.1060 dB MAE while improving RMSE from 0.2955 to 0.2503 dB. These support retained directional information, not rotation equivariance of the full network (Figure 2b).

3. Matched alternative inputs. We tested the suggested 50×50 raster FFT magnitudes in both predictors over seeds 0–2:

Predictor/inputMAE (dB)RMSE (dB)
Transolver-MS/SF0.11196 ± 0.001460.29960 ± 0.00638
Transolver-MS/FFT0.16113 ± 0.001840.39844 ± 0.00635
Point-DeepONet/SF0.11928 ± 0.002520.31503 ± 0.01076
Point-DeepONet/FFT0.15859 ± 0.001620.40464 ± 0.00610

Values are mean ± sample SD. SF lowers MAE by 30.5%/24.8%, respectively, and improves MAE/RMSE in all six pairs. Both arms use the same 21,000/4,500/4,500 split, geometry/descriptors, predictor capacity, 120 epochs and validation-MAE selection. Training-only standardized FFT vectors are padded from 2,500 to 4,032 entries; both inputs receive the same fixed orthonormal type-II IDCT and reshape to 7×9×64, without a learned adapter.

This compares complete pipelines, including sampling/preprocessing. The fixed raster has collisions for irregular templates; direct point-set SF evaluation avoids that rasterization step. We do not attribute the gap solely to angular compression. Table 2 and Appendix C.1 provide protocols and per-seed results.

4. Zero-spectrum MAE and lineage. Every zero arm starts from random initialization with zero standardized spectra and continues only from its own zero-spectrum checkpoints, including Point-DeepONet. Geometry, descriptors, spectral encoder and parameter count are retained. This tests input utility; the FFT comparison tests relative utility.

PredictorZero MAE, seeds 0/1/2 (dB)Zero mean ± SDFull mean ± SD
Transolver-MS0.137146 / 0.140267 / 0.1430950.14017 ± 0.002980.09181 ± 0.00190
Point-DeepONet0.140149 / 0.169558 / 0.1693430.15968 ± 0.016920.10558 ± 0.00423

All six full-spectrum pairs improve MAE/RMSE. Point uses base MSE training, then 40 MSE and 20 Huber epochs; Transolver continues its own formal checkpoints for up to 40 epochs. These staged comparisons retain different early-stopping trajectories. Seed-0 zero Transolver selects its initial checkpoint because no new epoch improves validation MAE, explaining its unchanged test MAE. Table C2 documents each lineage and selection; Table B3 gives schedules. The separate SF/FFT and holdout studies provide fixed-budget evidence.

5. Physical specification. Appendix B.1/Table B1 now consolidate the verified COMSOL/generator parameters: liquid water; array-level sound speed 1481.4 m/s, density 998.20 kg/m³; circular PMUT diaphragm radius 50 μm (diameter 100 μm); a reduced 4 μm silicon + 1 μm PZT-4 stack; square-template pitch 120 μm; equal in-phase 1 V drive. FEM-calibrated impedances/directivity determine coupled velocities. The target averages pressure magnitude over a centered 10×10 grid at 50 mm, at 23 frequencies. Dataset generation, splits and normalization are also specified there.

Reviewer_HpQ5OpenReview ↗

Thanks for the answers

while authors clarified some aspects of the paper (within the questions), and thanks for that, presenting/proposing a particular representation and its demonstrated benefits does not seem sufficiently novel enough to me. Authors basically say that they are not claiming to have invented the structure factor. They propose a particular way of organizing this classical information into an input that neural networks can use effectively. Hence, the scope and innovation is really limited. Useful..but not limited. My assessment about informativeness, soundness are improved. Presentation is better. but innovation is limited.

AuthorsOpenReview ↗

Thank you for acknowledging the improvements

Our statement about the classical structure factor was intended to credit the physical foundation. The learning question we address is how to organize this information into coordinates that different predictors can use and share.

The construction follows a specific physical requirement. Radial averaging makes rotated layouts indistinguishable, although a directional receiver can measure different responses. We therefore retain angular information alongside radial scales, producing a common coordinate system for global spatial correlations across different point layouts. This gives the input channels a consistent physical meaning across predictor architectures.

Your suggested FFT comparison helped us test whether this construction matters in practice. Across three seeds, it reduces mean MAE by 30.5% and 24.8% in Transolver-MS and Point-DeepONet relative to the tested raster-FFT pipeline under matched training budgets (Table 2).

We further found that the common coordinates support reuse of learned features. A spectral encoder trained in Transolver-MS can be transferred unchanged into Point-DeepONet while its geometry encoder and readout are trained separately. On the same acoustic task, this frozen transfer reduces mean MAE by 13.1% relative to training from scratch and by 35.5% relative to randomly initialized frozen features across three target seeds (Table 3). The latter comparison shows that useful information learned within one architecture remains usable within another.

We would appreciate your consideration of these findings on physically motivated representation design and learned-feature reuse in your final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.