{"id":"bb21ee35-f425-45f1-aaf8-2ac2fc9d0bf9","arxiv_id":"2504.15733","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A supervised neural operator maps pixelated images of 2D domains to Dirichlet eigenvalues and eigenfunctions, achieving about 1% relative eigenvalue error on held-out random shapes.","lead":"This paper trains convolutional and Fourier neural networks to take an image of a 2D domain and predict its Dirichlet eigenvalues and eigenfunctions. The approach scales and aligns domains first, and the paper reports about 1% relative eigenvalue errors on held-out random shapes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PCA main-axis alignment is ill-defined for shapes with degenerate or near-degenerate covariance, so the preprocessing is non-deterministic or discontinuous and the claimed generality to arbitrary 2D domains is unsupported.","rationale":"The reader's weakest assumption identifies exactly the PCA degeneracy in the main-axis alignment. This is the most load-bearing concern because it threatens the paper's implied claim of applicability to arbitrary 2D domains, not merely its reported metrics on the generated random-star-shaped training distribution. The paper's experiments use domains generated from random points with a minimum-distance constraint, which are almost surely free of exact symmetries; thus the reported accuracy does not test the preprocessing's behavior on disks, squares, or other symmetric domains. The concern is not an external consensus disagreement but an internal gap: the preprocessing is asserted to standardize orientation and simplify the task, yet for a class of inputs within the stated scope, the standardization is undefined or unstable. This can introduce artificial discontinuities in the input-output mapping, which the networks cannot faithfully represent. The issue is fixable, e.g., by detecting degenerate covariance and imposing a deterministic canonical orientation or by restricting the claims to the generated domain family. Because the empirical results on the tested distribution appear internally consistent and the concern is about generalization and preprocessing stability rather than a contradiction in the reported numbers, the conditional verdict remains appropriate. The reader's assessment and this stress-test agree on the core weakness; no verdict change is warranted.","tokens_in":12072,"tokens_out":10556,"duration_ms":104866,"concrete_test":"Construct a test set containing symmetric domains (disk, square, equilateral triangle, regular hexagon) and a one-parameter family of near-squares, e.g., rectangles with aspect ratio r ranging from 0.9 to 1.1 in small increments. Reproduce the preprocessing in Section 2.2 and record the principal-axis angle after alignment for each r, both with and without small pixelization noise. If the angle jumps by approximately 90 degrees near r=1 or is arbitrary for the symmetric cases, the preprocessing is discontinuous/non-deterministic. Then evaluate the trained 32x32+dp+ma CNN and FNO on these aligned images against FEM ground-truth eigenpairs (or retrain on an augmented dataset if model weights are unavailable).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as stated in the abstract and introduction, is that the method maps general 2D computational domains to their first few eigenpairs with high accuracy. A critical component of the pipeline is the main-axis alignment in Section 2.2, equations (2.6)-(2.8), which computes the covariance eigenvectors of boundary points and rotates the domain so its principal axes align with the coordinate axes, with an additional half-standard-deviation constraint to fix orientation.\n\nThis preprocessing is not well-defined for shapes whose covariance matrix has repeated eigenvalues: disks, squares, equilateral triangles, regular polygons, and any domain with sufficiently high symmetry. For such shapes, the eigenvectors are not unique; any orthonormal basis of the degenerate eigenspace is valid. The half-standard-deviation constraint does not disambiguate these cases because the left/right and lower/upper standard deviations are equal for all choices. Consequently, the same geometric domain can be presented to the network in different orientations depending on numerical rounding or small pixelization noise.\n\nMoreover, near the degeneracy set the mapping is discontinuous: for a near-square with covariance eigenvalues close to equal, a tiny perturbation can swap the ordering of the eigenvectors, producing a 90-degree rotation of the aligned image. Neural networks, being Lipschitz, cannot smoothly approximate a discontinuous input-output map. The paper does not analyze this degeneracy, does not include symmetric shapes in its generated dataset, and does not test the method on such domains. Since the paper claims generality to arbitrary 2D domains, this missing analysis leaves a load-bearing gap: for a natural and important class of domains, the method's input representation is not deterministic, so the claimed accuracy cannot be guaranteed and may fail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven operator surrogate for Dirichlet eigenvalue problems on two-dimensional domains. Domains are represented as pixel images and preprocessed by scaling, principal-axis alignment, and boundary-aware (averaged) pixelization. A CNN is used to predict eigenvalues and an FNO to predict eigenfunctions, with separate loss functions and training protocols. The authors report held-out test metrics on 20,000 randomly generated polygonal and smooth domains, including a first-eigenvalue MAPE of 1.15% and a first-eigenfunction RelL1 of 1.30% at 32x32 resolution, together with ablations of the preprocessing steps and results for the first 20 eigenpairs.","tokens_in":12380,"tokens_out":6791,"duration_ms":69989,"significance":"If the reported accuracy holds, the proposed pipeline is a useful fast surrogate for eigenvalue queries in shape optimization and inverse spectral problems. The paper has genuine strengths: it evaluates on a held-out test set rather than on training data, includes ablations of the preprocessing modules, and reports multiple error metrics. The main limitations are the unaddressed degeneracy of the main-axis alignment for symmetric or nearly symmetric domains, the missing specification of the network output dimensions across eigenvalue indices, and the absence of baselines and runtime comparisons. These gaps affect the reproducibility and the strength of the generality claim, but they are fixable within the scope of the manuscript.","major_comments":[{"comment":"The main-axis alignment is not well-defined for domains whose boundary-point covariance matrix has repeated eigenvalues (squares, disks, equilateral triangles, regular polygons). In those cases the eigenvectors in Eq. (2.7) are only determined up to an arbitrary orthogonal transformation of the degenerate eigenspace, and the sign-fixing rule based on left/right and lower/upper standard deviations does not break the tie because those standard deviations are equal by symmetry. As a result, the same geometric domain can be presented to the networks in different orientations depending on the eigensolver and on pixelization noise. Near the degeneracy set the aligned representation is discontinuous: an arbitrarily small perturbation can swap the ordering of the two covariance eigenvalues and rotate the aligned image by 90 degrees. Since the manuscript claims generality to arbitrary 2D domains and does not analyze this degeneracy, this is a load-bearing gap. Please either restrict the claims to domains with distinct covariance eigenvalues, introduce a deterministic and continuous orientation rule (e.g., based on higher-order moments), or add an explicit numerical test on symmetric and near-symmetric shapes showing that the predictions are stable.","section":"Section 2.2, Eqs. (2.6)-(2.8)"},{"comment":"The manuscript does not specify whether one CNN predicts all eigenvalues at once (e.g., an output dimension of 20) or a separate CNN is trained for each eigenvalue index. The same ambiguity applies to the FNO: it is unclear whether one network outputs a single eigenfunction and is retrained for each index, or whether several eigenfunctions are output together. This information is necessary to reproduce the results and to interpret Table 2, Figure 8, and Figure 12, which report per-index metrics. Please state the output dimensionality and whether parameters are shared across eigenvalue indices for both the CNN and the FNO.","section":"Section 3.1 and Section 4.1"},{"comment":"The statement that 'detailed pixelization and main axis alignment significantly improve the model accuracy' is only partially supported by the reported numbers. For the 32x32 setting, MAPE is 1.67% without any preprocessing and 1.75% with main-axis alignment alone, so alignment alone does not improve accuracy; the improvement to 1.15% comes from the combination dp+ma. Please revise the wording to describe the joint effect accurately, or discuss why alignment alone is not beneficial at 32x32.","section":"Section 4.1, Table 1"},{"comment":"No comparison is made with alternative operator-learning methods (such as DeepONet, a standard FNO used directly for eigenvalues, or a CNN-only baseline) and no runtime measurements are reported. The introduction and conclusion claim that classical methods are 'computationally intensive' and that the proposed method is efficient, but no timing comparison is given. Please add at least wall-clock times for the finite-element data generation versus online CNN/FNO prediction, and, ideally, a baseline comparison with a generic neural operator to calibrate the reported accuracy.","section":"Sections 3 and 4"}],"minor_comments":[{"comment":"The eigendecomposition is written as 'M = DM·vM', which is not a valid matrix factorization and is dimensionally inconsistent; it should be M = V D V^T, where V is the matrix of eigenvectors. Please correct this equation.","section":"Eq. (2.7)"},{"comment":"The loss function in Eq. (3.18) is the mean absolute error, but the surrounding text says 'the MSE penalizes large deviations' and no mean squared error is defined. Please make the formula and the description consistent.","section":"Eq. (3.18)"},{"comment":"The experiments are run without reporting variance across random seeds or across different train/test splits. Since the dataset is randomly generated, a single split may not be representative; please report means and standard deviations over at least three independent training runs.","section":"Section 4, statistical robustness"},{"comment":"The index in the weighted average θ*_{i,i+1} = w·θ_{i-1,i} + (1-w)·θ_{i,i+1} uses θ_{0,1} when i=1, which is not defined unless the vertex index is cyclic; please clarify the indexing convention for the vertex list.","section":"Section 2.1, Eq. (2.3)"},{"comment":"No code or data availability statement is included, and the finite-element solver used to generate the dataset is only referenced by a book; details such as the number of mesh elements and the treatment of high-order eigenfunctions would help readers reproduce the experiments.","section":"Reproducibility"},{"comment":"There are several typographical errors, including 'dateset' in Section 2.2, 'folowing' in Eq. (3.12), and the broken title 'OPERA TOR INFERENCE'; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical demonstration with moderate novelty. The PCA-alignment degeneracy is the most substantive technical concern and should be addressed directly; the missing output-dimension specification is a reproducibility issue that must be fixed. The absence of baselines and runtime data weakens the efficiency claims but is secondary. I do not see a load-bearing error that would require rejection; the issues can be resolved in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward piece of applied operator learning: it takes pixelated 2D domains as input and trains a CNN to predict the first 20 Dirichlet eigenvalues and an FNO to predict the corresponding eigenfunctions. On the author-generated test set of 4,000 random polygonal and smooth domains, the first eigenvalue MAPE is 1.15% (32x32) and the first eigenfunction RelL1 is 1.3%. Those are decent numbers, and the ablation tables show their preprocessing—particularly the boundary-aware pixelization—is pulling real weight.\n\nWhat's new: as far as the citations go, this is the first domain-to-eigenpair operator learning formulation. The divide-and-conquer architecture (CNN for scalars, FNO for functions) is sensible. The paper does not oversell the mathematics; it's a data-driven surrogate, and the held-out split is the right way to evaluate it.\n\nWhere I would push back, in increasing order of seriousness.\n\nMinor: equation (3.18) defines an L1 loss but the sentence below says \"MSE penalizes large deviations.\" Harmless typo.\n\nReproducibility: no code or data. For a paper built entirely on numerical experiments, that's a significant gap. Also, all tables report a single run; no seed variance. I wouldn't be surprised if the 1.15% MAPE varies by 0.2% across seeds. The authors should report mean and std over 3-5 runs.\n\nBaselines: The ablations compare against removing their own modules, but not against any alternative surrogate—say, a single CNN for both eigenvalues and eigenfunctions, or a regression on shape descriptors, or an FNO for eigenvalues as well. Without a baseline, \"high accuracy\" is only relative to what a trivial predictor does.\n\nThe larger issue is the main-axis alignment. The stress-test note is right: for shapes with degenerate covariance (squares, disks, equilateral triangles), the principal axes are arbitrary, and for shapes near that set a tiny perturbation can swap the ordering and rotate the input by 90 degrees. The preprocessing map then isn't continuous at that set. Since the paper claims generality to \"arbitrary 2D domains,\" this is an unsupported overreach. That said, the training and test sets are random polygons and Bezier shapes, which likely sit away from the degeneracy, so the reported numbers stand for that distribution. The paper should either restrict its claim to generic shapes or add some handling for near-symmetric domains.\n\nBottom line: this is a useful contribution for people building fast surrogates for spectral problems, shape optimization, or inverse queries. It deserves a serious referee. I'd send it out, with the expectation that the authors add code/data, seed variance, at least one meaningful baseline, and a discussion of the PCA degeneracy.\n\nBest,","headline":"A useful neural surrogate for Dirichlet eigenpairs on random 2D domains; solid experiments but with a real degeneracy in the PCA alignment step and missing baseline/reproducibility support.","tokens_in":12954,"tokens_out":3783,"would_cite":true,"duration_ms":34663,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35J15","65N25","68T07","65T50"],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural nets map any 2D shape to its first 20 eigenpairs.","keywords":["operator inference","elliptic eigenvalue problems","Dirichlet eigenvalues","eigenfunction prediction","Fourier Neural Operator","convolutional neural network","main-axis alignment","detailed pixelization"],"falsifier":"Take a square or disk, a domain whose boundary-point covariance matrix has equal eigenvalues, and run the pipeline twice on the same geometric shape, once with the principal-axis eigenvectors in one numerical orientation and once with them swapped by a tiny perturbation; if the predicted first eigenvalue differs by more than the reported MAPE, the deterministic orientation assumption fails. A second check is to compute test MAPE on a set of shapes with near-degenerate covariance and compare it with the full-set MAPE of 1.15%.","tokens_in":11892,"feed_emoji":"🧮","tokens_out":7086,"duration_ms":62724,"temperature":0.7,"pith_summary":"The paper tries to establish that the mapping from an arbitrary two-dimensional domain to its first few Dirichlet eigenpairs can be learned end to end from pixel images. It splits the learning task into two: a convolutional network maps the pixelated domain image to eigenvalues, and a Fourier Neural Operator maps the same image to eigenfunctions. A preprocessing module scales each domain to the unit square, refines boundary pixels with average pooling, and aligns each domain along its principal axes, so the networks see only intrinsic geometry. On 20,000 generated polygonal and smooth domains, the first eigenvalue is predicted to 1.15% mean absolute percentage error and the first eigenfunction to 1.30% relative L1 error at $32\\times 32$ resolution, with errors growing for higher modes. If correct, the method is a fast surrogate for repeated eigenvalue queries in shape optimization and inverse spectral problems.","feed_headline":"Neural nets map any 2D shape to its first 20 eigenpairs","feed_subtitle":"A CNN predicts eigenvalues and a Fourier network predicts eigenfunctions, with ~1% error on the first mode.","key_machinery":"The central mechanism is the image-based operator representation plus a three-part preprocessing module. Main-axis alignment computes the covariance matrix of boundary sample points, rotates the domain so its principal axes align with the coordinate axes, and resolves the two reflection ambiguities by comparing standard deviations across the left/right and lower/upper halves; this is what turns a rotation-invariant learning problem into a fixed-orientation regression. Detailed pixelization generates a high-resolution mask and average-pools it down to the target grid, so boundary pixels take fractional values that encode the area fraction of the domain inside the pixel; this is what prevents four narrow rectangles from sharing one input image. The eigenvalue CNN uses convolutional blocks with skip connections and a fully connected head, while the eigenfunction FNO uses four Fourier blocks with 16 modes and a masked, sign-aware, normalized L2 loss.","core_discovery":"The paper's central assertion is that a two-network operator pipeline can directly approximate the Dirichlet eigenvalue maps $\\Omega \\mapsto (\\lambda_1,u_1,\\dots,\\lambda_{20},u_{20})$ for simply connected Lipschitz domains in the plane, after the domain is represented as a preprocessed image. The authors argue that the decisive ingredient is not the network architecture alone but the preprocessing: scaling enforces the known $k^{-2}$ scaling law of Dirichlet eigenvalues, detailed pixelization keeps distinct narrow rectangles from collapsing to the same binary image, and main-axis alignment removes rotational and flipping ambiguity before training. With these ingredients, the CNN reaches $R^2=0.9988$ and MAPE $1.15\\%$ for the first eigenvalue at $32\\times 32$, and the FNO reaches PSNR $52.00$ and RelL1 $1.30\\%$ for the first eigenfunction; higher eigenfunctions degrade as their oscillation count grows. The paper further shows that relative eigenvalue error stays roughly constant across the first 20 eigenvalues, so the surrogate is not limited to the ground state.","pith_inferences":["An immediate stress test would be to restrict training to shapes with degenerate or near-degenerate covariance, such as squares or near-circles; if the principal-axis alignment is non-unique there, the model should either fail or require orientation augmentation to recover.","The same two-network split should transfer to Neumann or Robin boundary conditions, since only the labeled data and the mask change, not the architecture.","Predicted eigenfunctions could be recycled as initial guesses for a few finite-element refinement iterations, potentially combining the surrogate's speed with classical convergence guarantees.","Because the paper's largest-error samples are elongated domains, a test distribution with more extreme aspect ratios is likely to push MAPE above the reported 1.15%."],"forward_implications":["A trained model answers a new shape query with one forward pass, so repeated eigenvalue evaluations in loops such as shape optimization or real-time simulation no longer require a fresh finite-element solve per candidate.","The same trained networks generalize to unseen polygonal and smooth domains drawn from the same generator, giving a checkable baseline for operator-learning surrogates of spectral problems.","Relative eigenvalue error stays nearly flat across the first 20 eigenvalues, so a single model covers a spectrum rather than only the fundamental mode.","Eigenfunction accuracy falls off with mode number, so the practical envelope is the low- to mid-frequency range; higher modes are captured structurally but not quantitatively."],"supporting_citations":[{"why":"Supplies the linear Lagrange finite-element eigenvalue solver used to generate the ground-truth eigenpairs for training and testing.","marker":"[19]"},{"why":"Provides the Fourier Neural Operator architecture that the eigenfunction predictor is built from.","marker":"[14]"},{"why":"Provides the convolutional-network design (convolution, pooling, weight sharing) that the eigenvalue predictor adapts.","marker":"[12]"},{"why":"Supplies the random Bezier-curve scheme used to generate the smooth domains in the dataset.","marker":"[21]"}],"fun_headline_variants":["Two-net operator inference predicts eigenpairs for any 2D shape","Preprocessing key to neural net eigenvalue surrogates","CNN plus FNO maps domains to first 20 eigenpairs","Operator inference for elliptic eigenproblems via images","Scaling-aware neural nets nail Dirichlet eigenvalues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that principal-axis alignment gives every input domain a unique, stable orientation; for shapes whose covariance eigenvectors are degenerate (squares, disks, equilateral triangles), the orientation is arbitrary, so the same shape can be fed to the network in different rotations and the deterministic input-output mapping breaks.","fun_headline_variants_meta":{"raw":{"variants":["Two-net operator inference predicts eigenpairs for any 2D shape","Preprocessing key to neural net eigenvalue surrogates","CNN plus FNO maps domains to first 20 eigenpairs","Operator inference for elliptic eigenproblems via images","Scaling-aware neural nets nail Dirichlet eigenvalues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000406,"raw_usage":{"total_tokens":2101,"prompt_tokens":927,"completion_tokens":1174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1094}},"tokens_in":543,"tokens_out":1174,"duration_ms":9145,"temperature":1.0,"reasoning_tokens":1094,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:18:28.342950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a square or disk, a domain whose boundary-point covariance matrix has equal eigenvalues, and run the pipeline twice on the same geometric shape, once with the principal-axis eigenvectors in one numerical orientation and once with them swapped by a tiny perturbation; if the predicted first eigenvalue differs by more than the reported MAPE, the deterministic orientation assumption fails. A second check is to compute test MAPE on a set of shapes with near-degenerate covariance and compare it with the full-set MAPE of 1.15%.","supporting_citations":[{"cited_title":"Sun and A","cited_arxiv_id":null,"evidence_quote":"Supplies the linear Lagrange finite-element eigenvalue solver used to generate the ground-truth eigenpairs for training and testing."},{"cited_title":"A supervised neural network for drag prediction of arbitrary 2D shapes in low Reynolds number flows","cited_arxiv_id":"1907.05090","evidence_quote":"Supplies the random Bezier-curve scheme used to generate the smooth domains in the dataset."}],"review_version":1}