acceptodds
Under review as a conference paper at ICLR 2027

Physical Information, Numerical Resolution, and Learned Approximation in Born Scattering

Abstract

Learning wave-scattering operators presents three main challenges: resolving transmitted information, controlling discretization error, and approximating the operator from data. Using quadrature-weighted finite-aperture Born operators and DeepONets, we analyze their interplay. Operator spectra across frequency bands show a plateau and a steep tail, with a shifting knee that complicates global power-law fitting. Effective rank identifies key transmission directions. Receiver refinement shows first- or second-order convergence (rectangle or trapezoidal quadrature), making receiver requirements dependent on the integration rule and accuracy, while aperture extension affects physical coverage. In two-band studies, linear branches outperform MLPs; with width 64, wavelength matching cuts high-band linear-model errors by 66% and 52%. Refitting coefficients in learned bases remove most in-distribution error, but training-based POD with ridge prediction is more accurate at equal data, revealing an approximation gap. Full-wave checks find Born discrepancies of 4.5%–10.4% at nominal contrast, quantifying model accuracy but not neural full-wave generalization. Overall, we clarify the roles of physics, numerics, and learning, showing accuracy is distribution-dependent and not universally set by effective rank.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Born scattering operators, spectral complexity, operator learning.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Zae4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper discusses physical information, numerical resolution, and learned approximation in Born Scattering. The paper combines quadrature-weighted, matrix-free spectral experiments with empirical POD and trained-network comparisons. The methodology consists of three parts, namely, weighted acquisition operator, physical dimension and effective rank, and spectral estimation and controlled diagnostics. This work is organized by five research questions and three principal metrics. Experiments are then conducted to address each of the question.

Critical strengths
  1. This paper presents an interesting perspective that utilizing deep learning for electromagnetic purposes.

  2. The numerical results and discussions are organized according to the research questions, which is helpful and easy to follow.

Critical weaknesses
  1. The paper lacks comparison schemes against state-of-the-art. Hence, it is difficult to see the improvement of the advantages over existing methods.

  2. The paper would benefit from a clearer articulation of its core contributions. At present, the highlighted points do not sufficiently emphasize the novelty or impact of the approach. The related work section provides the position of this paper. However, the importance of filling those gaps is not provided. Lastly, incorporating a dedicated discussion section to contextualize the findings relative to existing literature would help clarify the work’s unique value.

Questions for authors
  1. How is the upper/lower-bound analysis? This is important to assess how close the approximation to the bound is, and to judge whether an approximation is required or not.

  2. How is the scalability of the framework? What happen if we change the simulation parameters? More importantly, why those values are considered for the simulation?

  3. How is the computational complexity of the approximation?

  4. Regarding the linear- and MLP-branch models, what are their parameters? How is the training hyperparameters? What happen if we change it to a bigger/smaller model?

  5. Could the authors clarify the rationale behind representing AA using its real and imaginary parts rather than its magnitude and phase?

Reviewer_r8KgOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studied the "finite-aperture, band-integrated Born operators mapping small medium perturbations to scattered acoustic measurements", mostly how the frequency band change the property of the Born output. It tries to separate three sources of difficulty: (i) physical information (how many directions the acquisition actually transmits), (ii) numerical discretization (receiver count and quadrature rule), and (iii) learned approximation (DeepONet-style models).

Critical strengths

The author is trying to analysis multiple factors and properties of the wave-scattering Born operator. From the best that can be guessed, the author is trying to analysis the information channel capacity of the operator map, the receiver resolution of the operator map, and the network learnability.

Critical weaknesses

Presentation. The paper is dense, with long sentences, jargons, and incomplete sentence logic. The main takeaways are hard to understand and follow. Several "separate" experiments (ranks, receivers, learning, full-wave) are loosely connected.

Limited scope and novelty of the core claims. Fast spectral decay of Born operators, the dependence of rank on frequency/aperture, and first/second-order quadrature convergence are largely expected from classical diffraction tomography and numerical analysis. The paper's contribution is mostly a bunch of random measurement, not a new method or theory.

The central questions are not actually answered. The paper answer its key research question ends with "cannot be tested." (line 416)

Questions for authors

Please elaborate more on the major motivation, key novelty, methodology, key research questions, and the major conclusions using more professional English with clear logic.

From the abstract: " Learning wave-scattering operators presents three main challenges: resolving transmitted information, controlling discretization error, and approximating the operator from data. " - How does the analysis solve the challenges? "Using quadrature-weighted finite-aperture Born operators and DeepONets, we analyze their interplay" - Explain the "interplay", as your method is not based on DeepONet. "In two-band studies, "linear branches outperform MLPs; with width 64, wavelength matching cuts high- band linear-model errors by 66% and 52%." - which linear branch are you referring to? “Effective rank identifies key transmission directions.” N_eff is a scalar count at a 0.1 threshold. How does a count identify directions? Did you examine the singular vectors? “Refitting coefficients in learned bases removes most in-distribution error.” Appendix F reports at least 99.17% of coefficient squared error removed. Why “most”? Is the remaining error span error, which refitting cannot fix? “…revealing an approximation gap.” The body calls it a “learned-span gap.” Are these the same thing? What is the evidence that it is an approximation gap and not an optimization failure? “…showing accuracy is distribution-dependent and not universally set by effective rank.” The normalized ranges are disjoint, so you say collapse cannot be tested. Which experiment supports “not universally set by effective rank”? “Full-wave checks find Born discrepancies of 4.5%–10.4%…” How does this connect to the learning results? Were any trained models evaluated on full-wave targets?

More questions: introduction and methods 9. “Does changing the frequency band alter the singular-value decay law, or primarily the number of strongly transmitted directions?” Can you answer this in one sentence? Table 1 shows the global exponent falling from 7.04 to 1.39, so which interpretation do you conclude? 10. “A widening plateau followed by a steep tail can require a larger approximation rank without a slower asymptotic decay law.” Which figure or table shows this directly? 11. Sec. 3.2: γ “is not assumed to be a coverage fraction.” Then what does γ mean? N_eff/N_0 drops from 0.48 to 0.37, so is the unit-slope hypothesis rejected, and what is the value of N_0 as a predictor? 12. N_eff is 143.8 for 15–30 Hz in Table 1, but 98.1 (real-stacked, coarser grid) in Appendix B. Which value should a reader use, and how should the two be compared?

More questions: experiments 13. Sec. 5.4 says the network is within 0.01 of the empirical POD reference for p ≤ 32. Appendix F shows training-POD+ridge errors of about 1e-6 against about 1e-2 for the networks. Why do the two studies disagree? Is it the discretization, the training protocol, or the metric? 14. Sec. 5.3: the fit uses only {33, 49, 65, 97} receivers. Why that subset? Left-rectangle on shifted arrays never reaches 1% by 97 receivers. What does this imply for practitioners? 15. Sec. 5.5: illumination fractions of 0.85 and 0.32 are reported, yet “do not establish the cause of forward error.” What is the purpose of Fig. 3c? 16. Appendix G: 30 of 288 cases fail the relative grid-change gate, and the high-band OOD results are identical across ensemble labels. How many independent comparisons are there, and why only eight samples per condition? 17. Why a single seed and a 1,000-epoch budget? Can you report variance across seeds for the 66% and 52% claim? 18. “Which result should a reader take away if they remember only one?” Please state it in one or two sentences, and say which experiment supports it.

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

The paper uses the linearised (Born) operator for 2D acoustic scattering, which maps small changes in a medium to scattered waves recorded at surface receivers, as a test case for operator learning.

First, the paper studies the operator itself. Its singular values form a plateau followed by a steep drop, and the point where the drop begins (the knee) moves with the frequency band. The authors argue that fitting a single power law to the whole spectrum is misleading, because the moving knee can mimic a change in the decay exponent; local slopes vary by a factor of 60–130 across each spectrum. What matters is where the knee is, which the paper measures with an effective rank: the number of significant singular values. The effective rank grows roughly as a simple geometric count predicts. Adding receivers makes the computed spectrum converge at first or second order, depending on the integration rule.

Second, the paper trains small DeepONet-style networks to learn the operator, and compares them with projections onto the operator’s singular vectors and onto principal components of the outputs (POD). It splits the network’s error into the error from a poorly learned output basis and the error in the coefficients on that basis. Refitting the coefficients by ridge regression, keeping the learned basis fixed, removes most of the coefficient error. However, POD on the training outputs plus ridge regression is still much more accurate than the trained networks.

Critical strengths

The authors check their computations thoroughly.

The operator is well designed and analysed thoroughly.

Limitations are stated clearly

Critical weaknesses

The results are mostly already known. A plateau followed by a steep drop is classical behaviour for band-limited wave operators. Also, a linear map being learned best by linear methods (POD plus least squares) is expected.

The learning experiments use 800 training samples, and a fixed grid. The main research question is not answered as the network width is not compared to the spectrum.

Questions for authors

Can you choose decoder widths so that p/N_eff overlaps across bands, and then test RQ5?

Do any of the conclusions carry over to a nonlinear (full-wave) target or a heterogeneous background?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.