acceptodds
Under review as a conference paper at ICLR 2027

Neural networks for spectral optimization

Abstract

Given a functional dependent on the spectrum of a differential operator, we address the problem of finding a domain which optimizes this functional. PDE solvers might be used to tackle this optimization. It is however computationally expensive. We propose two neural network models which learn the spectrum directly from the geometry of the domain and can be used to optimize the domain from one or more eigenvalues. We investigate two representations. The first encodes the domain through Fourier coefficients and a light MLP, which is efficient on star-shaped geometries, achieving a precision of 0.2%. Through a rescaling of the coefficients the designed models satisfy the scaling law of the eigenvalues. Additionally, averaging the outputs of the trained surrogates over rotations and reflections induces invariance for these transformations. The second is a model that takes the landscape function, the indicator function and the gradient of the landscape function. A Gram-Schmidt process produces orthogonal eigenfunctions as output of the model along with the associated eigenvalues. The landscape model reaches 1% mean relative error on the first ten eigenvalues, compared with 4% for an FNO model. Replacing the landscape by an SDF worsened both prediction and optimization errors. The trained model also generalizes from synthetic shapes to domains given as classical image dataset. The resulting surrogates of both approaches recover classical spectral optima such as the disk for the first eigenvalue or the conjectured minima of higher eigenvalues. This confirms that our models produce accurate differentiable estimates of eigenvalues, which can be used in shape optimization problems involving spectral quantities.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spectral geometry, Neural surrogates, Geometric deep learning

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_Y7YzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes two differentiable neural surrogates for the first ten Dirichlet eigenvalues of planar domains, aimed at shape optimization. The first treats star-shaped domains: the boundary is a truncated Fourier series, and three small MLPs predict area-normalized eigenvalues, with rotation/reflection averaging at test time The second uses the torsion/landscape function as the input: a three-channel image of a smoothed indicator, the normalized landscape, and its squared gradient is passed to a U-Net whose decoder emits ten functions that are orthonormalized by a differentiable Gram--Schmidt step, with eigenvalue ratios read off Rayleigh quotients. The landscape model is the headline result, reaching about 1%1\% mean relative error over ten eigenvalues and outperforming author-reimplemented FNO (2.33%2.33\%) and DeepONet (1.47%1.47\%). The authors motivate the landscape representation with a stability inequality relating landscape functions to eigenvalue perturbations.

Critical strengths
  1. The landscape-as-learned-input representation is new, and the ablation is clean. Feeding the torsion/landscape function to a learned geometry-to-spectrum map is a new contribution. The same-backbone comparison shows that the input representation, not the architecture, is doing most of the work. This is the contribution the paper should be built around.
  2. The surrogate is amortized and differentiable, so it is actually usable for optimization. It predicts ten eigenvalues in one pass and exposes gradients for arbitrary functions of them. The case studies recover the disk and two-disk optima to within 1.43%1.43\% and 2.35%2.35\%, which is a credible check that the tool works.
Critical weaknesses
  1. The abstract does not match the paper's own tables. It says FNO is at 4%4\%, but Table 5 says 2.33%2.33\%. It says the Fourier model reaches 0.2%0.2\%, but the retained model reaches 0.237%0.237\% at λ10\lambda_{10}. It says replacing the landscape with an SDF worsened optimization, but no SDF optimization experiment was run. It calls the MPEG-7 error 4.83%4.83\%, which is the median; the mean is 6.88%6.88\%. The corrected numbers are still decent, which is exactly why this is damaging: it looks like the headline text was not checked against the results.

  2. The main theoretical justification is not actually established. The stability inequality is presented as "mathematically proved," but there is no citation, no proof, and no value or dependence for CC. The cited David et al. paper is about the landscape on Rd\mathbb{R}^d for random Schrodinger operators and does not contain this statement, and Arnold et al. (2019) describe the landscape-to-spectrum method as only partially justified. Since this inequality is the stated reason for using the landscape representation, the motivation is missing.

  3. Several technical claims do not match the implementation. Eq. (3) is said to ensure positivity of the border function, but summing the printed per-mode bound gives about 1.185a01.185 a_0, which is greater than a0a_0, so it does not imply ρ>0\rho > 0; the released code uses a total-amplitude budget instead. The contributions say the Fourier model uses translation invariance, while Section 3.1 says translations are explicitly not imposed. The "single forward pass for arbitrary raster domains" claim only holds for the landscape model; the Fourier model takes 30 coefficients and averages over 12 rotations and reflections for three networks.

  4. The novelty is narrower than the title suggests, and the reference list is thin and padded. Neural, differentiable shape optimization of spectral and geometric functionals has already been done by Martinet and Ftouhi (2026, arXiv:2602.14881), and An and del Castillo (2025, arXiv:2507.07073) already learn Laplace-Beltrami spectra from geometry. NEO (arXiv:2605.24390) already does amortized eigenspectrum prediction, and Rowan et al. (2025) already combine Rayleigh quotients with Gram-Schmidt. The paper's 17 references are thin for the claims and include tangential items that do not support the argument. The defensible contribution is the amortized Dirichlet-spectrum surrogate with a landscape input, and the paper should be framed around that.

Questions for authors
  1. The stability inequality is the stated reason for using the landscape representation, but it has no citation, no proof, and no value or dependence for CC. What is the precise theorem and reference, and what does CC depend on? The cited David et al. paper appears to be about the landscape on Rd\mathbb{R}^d for random Schrodinger operators, not this Dirichlet eigenvalue statement.

  2. Eq. (3) is said to ensure ρ>0\rho > 0, but summing the printed per-mode bound gives about 1.185a01.185 a_0, which is larger than a0a_0. Was the implemented constraint the per-mode bound or a total-amplitude budget, and what exactly was used? Also, why do the contributions claim translation invariance when Section 3.1 says translations are not imposed?

  3. How do you distinguish the contribution from Martinet and Ftouhi (2026, arXiv:2602.14881), An and del Castillo (2025, arXiv:2507.07073), NEO (arXiv:2605.24390), and Rowan et al. (2025), who already combine Rayleigh quotients with Gram-Schmidt?

  4. Tables 2 to 6 report no seed spread, and the optimization uses the best of ten starts without a success rate. Given that some headline gaps are small, for example landscape 0.93%0.93\% versus DeepONet 1.47%1.47\%, how do you establish that these differences are stable rather than run-to-run variation? Can you report per-seed errors and optimization success rates?

3 comments

We thank Reviewer Y7Yz for the constructive comments and questions. Before answering the questions, we want to clarify the problem we address in the paper is variational spectral shape optimization. In more detail, "the methods consist in deforming the domain Ω\Omega to minimize an objective functional F(λ1(Ω),…,λk(Ω))F(\lambda_1(\Omega), \dots, \lambda_k(\Omega)) under area constraint Antunes & Freitas (2012). This variational problem is different from spectral shape analysis in geometry processing Reuter et al. (2006); Yigit & Wilson (2025); Velich et al. (2026) or Kac’s inverse problem Kac (1966); Gordon et al. (1992), which either compute Laplace-Beltrami descriptors on surfaces for shape matching or reconstruct a shape from a target spectrum.” (please refer to the updated introduction).

Questions:

Question 1: What is the precise theorem and reference, and what does depend on?

We added the specific mathematical statement with the explicit constant and the precise references. The inequality, as now stated, is a direct consequence of Theorem 3.4, Bucur (2003) and Theorem 3.1, Cheng & Yang (2007).

Question 2: Was the implemented constraint the per-mode bound or a total-amplitude budget, and what exactly was used?

We removed the claim about translation invariance. We also corrected the paragraph describing the database.

Question 3: How do you distinguish the contribution from Martinet and Ftouhi (2026, arXiv:2602.14881), An and del Castillo (2025, arXiv:2507.07073), NEO (arXiv:2605.24390), and Rowan et al. (2025), who already combine Rayleigh quotients with Gram-Schmidt?

We hope we clarify our contribution in the contributions section with

“ Additionally, Martinet & Ftouhi (2026) used invertible neural networks as surrogates of the shapes in their exploration of Blaschke-Santaló diagrams rather than the functionals themselves, their work focuses on convex bodies parametrized by a transformation of the disk making use of analytic formulae for the shape functionals and their derivatives. ”

Moreover, following the point cloud representation in geometry processing (Marin et al., 2020; Pang et al., 2024; An & del Castillo, 2025); Yang et al. (2026), we trained a PointNet on the same dataset. Please refer the last paragraph before the conclusion. Briefly, PointNet performs well on simply connected and convex geometries (triangle 5.12%5.12\%, polygon 5.65%5.65\%) but fail for more topologically complex shape (several connected components 18.18%18.18\%, non convex 34.93%34.93\% compared to 0.97%0.97\% and 2.55%2.55\%).

Finally, Rowan et al. (2025) optimize a neural network through the Rayleigh quotient for a single fixed domain, requiring a full training whenever the geometry changes. Our methods predict the first eigenvalues and eigenfunctions in a single forward pass and are differentiable with respect to domain deformations.

Question 4: how do you establish that these differences are stable rather than run-to-run variation?

We thank the Reviewer for raising this point. To clarify, the dispersion in the updated Table 5 corresponds to intra test set standard deviation across the test set. Retraining completely every model on N=100000 domains was not feasible within the rebuttal window. Our model achieves not only a lower mean relative error (0.92% vs. 1.48%1.48\% for DeepONet and 2.42%2.42\% for FNO) but also has a low inter shape standard deviation (0.86%0.86\% vs. 1.61%1.61\% for DeepONet and 1.77%1.77\% for FNO).

I thank the authors for their rebuttal, which addressed some of my concerns.

Even though I believe that the idea is worth pursuing and the results are interesting, I will maintain my rejection recommendation, as I believe that the work could really benefit from a fundamental rewriting.

I would also advise the authors to submit the rebuttal earlier than 15 hours before the end of the discussion window.

We thank you for the follow-up comment and for acknowledging that the core idea is worth pursuing and the results are interesting.

We apologize for posting the revised manuscript late in the discussion period. Training the PointNet on our dataset alongside the new optimization experiments took significant compute time. Should the paper be accepted, we commit to revising the presentation and framing of our contribution for the camera-ready version.

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

The paper tackles the problem of computing the frequencies of planar shapes via a neural network applied to a representation of the input shape. The goal is to have an estimate that is robust to small perturbations of the shape itself, eg adding a boundary "noise" point inside the shape. The authors evaluate by supervising MLPs on starry closed simple planar curves, using a Fourier coefficient representation on a radial parametrisation of the curve, or other small networks on the voxelisation of the shape (image grid of planar shapes), using the landscape function.

Critical strengths
  • The work is basic and does not hide behind artificial layers of complexity
  • Learning Laplacian eigenvalues (and eigenfunctions) is something I find interesting, even if it is niche
Critical weaknesses
  • Unfortunately, I did not find the paper interesting. The lack of motivation, absent theory, and empty/weak/poorly performed experiments removed any interest I had in the work by the time I reached its end. I would encourage authors to rework their paper from start to finish: provide a solide motivation, whether for the theoretical or applied communities, introduce novel theoretical insights (if possible), perform clean experiments (compare with and provide baseline results, provide ablations on design choices, be more ambitious in their applications by exploring other types of shapes typically 3D shapes...). I also strongly encourage authors to lean more towards shape optimisation rather than eigenvalue estimation, by clearly defining the problem and possible solution(s) right after the related works in an easily readable and self-contained manner, and then reorient their experiments around this task.

  • The paper is not sufficiently motivated to justify its existence or usefulness. It does not provide any new theoretical insights for heavy practitioners already having tools to compute Laplacian spectra, nor does it provide a novel rich practicality that theoreticians can use. In fact, many of the assumptions to legitimate the paper are not true:

    • Lines 115-117: the authors argue that shape optimisation is a challenge for mathematicians working in spectral geometry as it requires some knowledge in a couple elementary fields (PDE, calculus of variations, optimisation, and knowing some code). Not only does this somewhat insult the intellect of "mathematicians working in spectral geometry", who if they work in spectral geometry need to know aspects of these listed fields, and their implementation capabilities in "scientific computing", which is just caricatural. I would even add that even for non mathematicians and neophytes in spectral geometry, shape optimisation is within reach thanks to the help of modern agentic coding and LLMs. This is 2026 after all...
    • The authors motivate this work by repeatedly positioning as a shape optimisation work (see title and first line of the abstract), ie the task is to find the shape having a given spectrum. This task is of intellectual interest and is a cute mathematical problem. However, this is not the problem actually studied by the authors, although it is incorrectly phrased as such. The problem studied is simply spectrum estimation given a shape, and not its dual more interesting problem of shape optimisatino given a spectrum.
    • The most influential line of though linking a shape and its spectrum is not even mentioned or referenced. The authors must discuss the question: "Can you hear the shape of a drum", not just "Why are all drums round?" (l36), and cite the original work [1], its negative theoretical answer [2], and its positive answer in practice [3]. The paper lacks to establish the proper foundations on which to build on.
  • The authors solely focus on eigenvalues. Eigenfunctions, ie Fourier functions of the domains, are not studied or searched for, yet are often more useful in practice than the eigenvalues: for instance in shape analysis, the Laplace-Beltrami operator (LBO) provides the essential eigenfunction basis to perform linear shape analysis on curved domains, eg linear matching via functional maps on spectral coefficients. The eigenfunctions are often not used. I strongly encourage the authors to incorporate eigenfunctions in their work. Note that other works, even recent ones [4], computing Laplacian spectra can also provide the eigenfunctions, thus revealing a major drawback of the authors' approach compared to existing works (which need to be cited).

  • There is no proper method in this work, nor theory to compensate. The empirical method is simply 101 supervised learning. There is no innovative contribution in the framework or network either. In fact, there are many existing works that learn Laplacian decompositions (eigenvalues only or both eigenvalues and eigenfunctions), in supervised or unsupervised ways [4,5,6,7]. The authors should acknowledge these (and more) works, discuss them, cite them, and clearly position themselves against them to justify why their approach is "better" (in some sense that the authors can find).

  • Experiments lack ambition. The shape analysis community has moved to 3D shapes (some even to high-dimensional manifolds of latent representations), eg Faust/Tosca... datasets, which are of higher interest as more complex and more naturally present different explicit representations for the same implicit geometry (non-rigid isometric deformations of shapes like a person moving). Half the experiments are on starry planar simple closed curves (that are smooth and low pass, l176), the other half connects such shapes. Theses are not very interesting shapes.

  • The arguments motivating the work, such as the stability to small perturbations of the shapes for instance due to boundary points inserted inside it, are not empirically studied. The authors should perform these experiments to confirm or infirm their motivation that could justify the work.

  • The experiments are of exceptionally poor quality. Excluding the fact that they are not of high interest and lack ambition, there are major fatal flaws in their reporting:

    • Math notations are mixed with number choices arbitrarily, eg 2N+1 (l176) becomes 15 (l179). There is no motivation, explanation, or ablation in these cases.
    • Only a couple design choices are ablated, and they relate to the architectures of the networks, but a lot more interesting ones, especially those related to the problem being solved (input representation/pixel size, number of eigenvalues...) have not been performed. But even the architectural designs are not sufficiently ablated and appear extremely odd. For instance l222, MLPs are of shape [30,192,192,192,10], which is unusual for MLPs and deserves some investigation on the hidden-with dimensions, or fancy SiLU activation (not cited by the way) is used but other much more popular fancy activations are not like GeLU.
    • Most experiments are not clearly described: how shape optimisation is done is literally absent (or at the very least not self-contained), how random shapes are generated is unknown (for example l194: "We generate [...] satisfying this assumption": this is so vague and so many options possible - nothing mentioned on the choice of randomness, nothing mentioned on how the decay bound is enforced...), and so on. There is no appendix to provide these technical details that are necessary in top venue papers in 2026.
    • The authors only aim to recover the first 10 eigenvalues. This is exceptionally short for any practical use. In 3D shape analysis, we usually truncate after a couple hundred for accurate approximation, and around 30-50 if speed/memory becomes an issue. Just 10 is too few to properly encode the geometry of complex shapes.
    • Stability and other perturbations analyses are not performed.
    • The very naive approach from the authors suffers from explicit representation, and the authors naively resolve to augmentation (reflection and a few rotations) to be (slightly) more robust to rigid transforms. Augmentation is known to not solve this problem, although it helps, and using "12 fixed rotation" angles (l228) only tackles the finite rotation group and not all possible rotations, why not instead take random rotation angles and learn to be invariant to the continuum rather than the discrete group? Equivariant/Invariant frameworks could/should also be explored to trivialise the issue.
    • The authors average results on several networks (l238). It is not clear why (simply claiming for stability l221 is not enough), and there is no ablation on this (again no ablation).
    • Only partial results are provided. Table 1 (which is not a table by the way) provides number for the 5-th to 10-th eigenvalue, what about the 1-st to 4-th? It is not mentioned in Table 1 what shape this is computed on, making this result absolutely void. Regarding qualitative results, only a handful of results are provided, without groundtruth or baselines for comparison (eg Figure 2) making them not interesting.
    • There is no comparison with baselines or existing works. While there are some mentions of similar results (especially to Antunes-Freitas), they are neither provided quantitatively nor qualitatively anywhere in the submission. This on its own is already sufficient for me to recommend the minimal score.
    • Failure cases are not shown, discussed, or studied. Neither are limitations.
    • Standard deviations are not reported (Table 5).
  • I question how sensitive the eigenvalue prediction algorithm is to non perfect shapes, typically non-watertight shapes where SDFs no longer exist (there is no inside or outside and thus no clear Ω\Omega to even define the landscape function).

  • Some results are claimed as known, yet would deserve to be cited as is good practice in writing research papers. For instance l142 "It has been mathematically proven" should be accompanied with a reference to where the proof first emerged or at the very least a widely accepted textbook including this result. In general the authors should cite whatever is mentioned, even coding libraries like SymPy [8] (l207).

  • Terms are not always defined properly (eg already in the abstract FNO l23 in the abstract is undefined).

  • There are quite a few typos. The biggest one is l145: there is probably a missing absolute value on the ω\omega difference, but other minor bugs can be found like l160: method predicts --> methods predict.

[1] Kac, Can one hear the shape of a drum?, AMM, 1966

[2] Gordon, One cannot hear the shape of a drum, Bulletin of the AMS, 1992

[3] Reuter et al., Laplace-Beltrami spectra as 'Shape-DNA' of surfaces and solids, CAD, 2006

[4] Velich et al., Learning Eigenstructures of Unstructured Data Manifolds, CVPR, 2026

[5] Yigit & Wilson, LBONet: Supervised spectral descriptors for shape analysis, PAMI, 2025

[6] An & Castillo, An AI Approach for Learning the Spectrum of the Laplace-Beltrami Operator, arxiv, 2025

[7] Pang et al., Neural laplacian operator for 3d point clouds, SIGGRAPH/ToG, 2024

[8] Meurer et al., SymPy: symbolic computing in Python, PeerJ CS, 2017

Questions for authors

If the authors believe that I have completely missed the point of their work, I first apologise, and second invite the authors to reformulate/reexplain what they have achieved and why it is of significant practical or intellectual interest, espeically when put into the context of existing works.

I would also like the authors to specifically address why they limited themselves to extremely simple (near) convex/starry planar shapes. Could their method tackle 3D shapes and how would that compare to existing methods?

If the authors wanted the work to be about shape optimisation, why didn't the authors clearly present this problem and how it is solved? Instead the method is focused on the estimation of eigenvalues, which is a different problem.

Beyond questions, the experimental methodology and presentation of results needs to be completely revisited and improved.

While my comments might be tough, I am a reviewer that is happy to change their grade given a proper convincing response to the review. Unfortunately, there are so many major fatal issues in this work that it seems extremely unlikely to me that the authors will manage to change my mind without completely rewriting the entire paper. I advise the authors to take this into account should they want to rebut.

2 comments

We thank Reviewer hCCT for the questions allowing us to clarify the problem we address in this paper.

Questions:

If the authors believe that I have completely missed the point of their work, I first apologise, and second invite the authors to reformulate/reexplain what they have achieved and why it is of significant practical or intellectual interest, espeically when put into the context of existing works.

Thank you for this question. After reading the review, we realize that our introduction does not relate geometry processing and shape optimization with variational spectral shape optimization in PDEs (Dirichlet, Neumann, ...) which is the problem addressed in our paper. We think we clarified this in the introduction with:

“ Spectral shape optimization consists in searching geometries satisfying extremal properties of various spectral functionals. The methods consist in deforming the domain Ω\Omega to minimize an objective functional F(λ1(Ω),…,λk(Ω))F(\lambda_1(\Omega), \dots, \lambda_k(\Omega)) under area constraint Antunes & Freitas (2012). This variational problem is different from spectral shape analysis in geometry processing Reuter et al. (2006); Yigit & Wilson (2025); Velich et al. (2026); Yang et al. (2026) or Kac’s inverse problem Kac (1966); Gordon et al. (1992), which either compute Laplace-Beltrami descriptors on surfaces for shape matching or reconstruct a shape from a target spectrum. Furthermore, spectral optimization is important for open mathematical conjectures Antunes & Freitas (2012); Bogosel (2023); Martinet & Ftouhi (2026) and has many applications including inverse engineering design like vibrational bandgaps in acoustic metamaterials Khazi et al. (2026); Sigmund & Jensen (2003). To perform efficiently a spectral optimization process, one must be able to evaluate the spectrum for any domain. ”

We also stress that we focus on the 2d case as it is the classical numerical spectral shape optimization setting.

Table 1 (which is not a table by the way) provides number for the 5-th to 10-th eigenvalue, what about the 1-st to 4-th?

In Table 1 we reference the results obtained by Antunes and Freitas for the optimization with area constraint of the 5-th to 10-th eigenvalue. They do not give estimates for the 1-st to 4-th because they are either proved or conjectured (with strong numerical evidence) to be minimized by the disk (for the 1-st and the 3-rd) or a union of two disjoint disks (for the 2-nd and the 4-th), the values being given by zeros of Bessel functions.

how sensitive the eigenvalue prediction algorithm is to non perfect shapes, typically non-watertight shapes where SDFs no longer exist. The arguments motivating the work, such as the stability to small perturbations of the shapes for instance due to boundary points inserted inside it

In contrast to the SDF, the landscape function solves the PDE

−ΔwΩ=1 in Ω,wΩ=0 on ∂Ω.-\Delta w_\Omega=1 \text{ in } \Omega, w_\Omega=0 \text{ on } \partial \Omega.

directly from a grid with occupancy mask 1Ω1_\Omega without boundary contour extraction, and remains continuous under capacity-vanishing internal perturbations. We added experiments where we optimize eigenvalues to illustrate the robustness of the landscape function to holes (small and big). Please refer to Figure 6.

Beyond questions, the experimental methodology and presentation of results needs to be completely revisited and improved.

We added more experiments. To highlight the fact that we use eigenfunctions, we updated Figure 5 with the predicted eigenfunctions and the domain for each gradient descent iteration along with the eigenvalue relative error against FreeFem.

Following the point cloud representation in geometry processing (Marin et al., 2020; Pang et al., 2024; An & del Castillo, 2025), we trained a PointNet on the same dataset. PointNet performs well on simply connected and convex geometries (triangle 5.12%5.12\%, polygon 5.65%5.65\%) but fails for more topologically complex shape (several connected components 18.18%18.18\%, non convex 34.93%34.93\% compared to 0.97%0.97\% and 2.55%2.55\%).

We added standard deviations across test domains in Table 5.

I thank the authors for their response.

As I mentioned in my original review, it would be hard to change my mind regarding this work without substantially rewriting the (whole) submission. This seems to be an issue somewhat shared with Reviewer Y7Yz: "as I believe that the work could really benefit from a fundamental rewriting".

While I acknowledge that the authors have made some efforts to address my concerns, they still remain mostly as is. I still think this work is not at its core about shape optimisation, as the authors desire it to be, but it is on calculating eigenvalues. This is actually explicit from the last sentence of the new motivation paragraph: "To perform efficiently a spectral optimization process, one must be able to evaluate the spectrum for any domain". The work, as it is written, makes shape optimisation as the downstream application on which to evaluate the spectrum predictor, which is the core aspect of the work. In fact the actual shape optimisation part is still not self-contained in the revised manuscript, making it unclear how it is actually done by reading the submission alone, nor does there seem to be any fundamental contribution on the actual shape optimisation part (given a spectrum extractor).

Nevertheless, I appreciate the effort made by the authors, such as going towards 3D experiments, looking at the various pointers I shared with them, and taking some initiatives further (such as the relevant Marin et al. 2020). There probably is something valuable in this work, and the actual shape optimisation problem is intellectually cute regardless how niche it is. But in the current state, I just cannot recommend the work for acceptance. I strongly advise the authors to take the time needed to heavily rewrite the work, perform the additional experiments, and create a much stronger polished submission to be reviewed from scratch in the future. I would personally love to see the authors pull it off.

I have thus decided to maintain my score, although I would not mind giving an additional half a point to it for the authors' efforts.

PS: The authors included the references in my review in their revised manuscript. However, Velich et al. is cited as "Velich et al.". I am pretty sure Velich's coauthors, namely Piven, Bensaïd, Cremers, Dagès, and Kimmel, would love to get their credit too! This is the only reference with this typo. I advise authors to double check and clean up the references if needed.

Reviewer_F8U3OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Note: The paper falls far outside my area of expertise. Since I am unable to fairly judge this paper, this is a placeholder review, and the AC was notified.

The paper addresses spectral shape optimization: among all 2D shapes of a fixed area, find the one that minimizes some function of its eigenvalues, i.e: the shape with the lowest first eigenvalue is the disk. Normally every optimization step requires an expensive solver. Instead, the authors train neural networks that predict a shape's first 10 eigenvalues, so the shape can be optimized by gradient descent through the network.

The paper proposes two models:

  1. Star-shaped domains are parametrized by 31 Fourier coefficients of the boundary radius. They are fed to an ensemble of small MLPs that predict eigenvalues (multiplied by the shape area), and averaged over 12 rotations and reflections (as rotating or reflecting a shape does not change its eigenvalues).

  2. A topological aware model that handles holes and multiple parts. Here, a U-Net takes a 64x64 image of 3 channels: the landscape function - a smooth map computed from the shape, its gradient and an indicator function for the shape's mask. The output are approximated eigenfunctions from which eigenvalues are computed.

The models are compared against the baselines of FNO and DeepONet, mask and SDF inputs, and tested on the MPEG-7 dataset of real silhouettes. In optimization tests, the proposed models are able to recover known answers, such as the disk for the first eigenvalue and two equal disks for the second eigenvalue.

Critical strengths

The Unet model finds the best known optimal shapes in simple tests. The model for star shaped domains comes close. Code and weights are released.

Critical weaknesses

The Unet model requires a landscape function as input. In that case, a solver is still required to compute it?

Questions for authors

n/a

1 comment

We thank Reviewer F8U3 for their positive evaluation and for highlighting our spectral shape optimization results and code release.

Question:

"The U-Net model requires a landscape function as input. In that case, a solver is still required to compute it?"

Thank you for raising this question. We have now clarified in Section 4.2 of the revised paper. Computing the landscape function wΩθw_{\Omega_\theta} does not require running a computationally expensive FEM mesh generator. Either the landscape function is known analytically in closed form or it is obtained thanks to finite difference solving which is computationally efficient and differentiable.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General rebuttal comment

We would like to express our sincere thanks for the time and efforts made by the reviewers to provide thoughtful and constructive questions about our paper. In response of the reviewers, we uploaded a revised version of our paper with the corrections highlighted in purple. Please find in answer of each individual review a summary of the improvements.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.