{"id":"9fd3e765-66fc-4aa1-9c31-a89158feea58","arxiv_id":"2411.15828","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tensor neural network method (FieldTNN) computes Maxwell cavity eigenvalues in tensor and non-tensor domains with a divergence-free penalty that suppresses spurious eigenpairs.","lead":"This paper introduces FieldTNN, a tensor neural network for vector fields, and applies it to Maxwell eigenvalue problems in square, L-shaped, and inhomogeneous cavities in 2D and 3D. The method combines a Galerkin formulation with a divergence-free penalty to compute accurate eigenvalues while filtering spurious modes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The divergence-free filter is asserted rather than demonstrated: under the paper's own H(div)/H(curl) norms, a real eigenpair has ratio 1/(1+λ), not ~0, and Table 6.1's near-zero H(div) values imply a different seminorm is being used.","rationale":"The reader identified the filter-ratio heuristic and the degenerate-spurious-mode gap as the weakest assumption. I agree this is the central risk, and I add a sharper, more elementary objection: under the norms defined in Section 2, the stated rationale for the ratio is internally inconsistent. For any normalized divergence-free eigenpair, the full H(div) norm contains ‖E‖²_L2 and does not approach zero, and the ratio equals 1/(1+λ), which is not close to zero for low eigenvalues. The numerical tables showing |E_NN|_H(div) ~1e-9 are only consistent with using the seminorm ‖divE‖, so the paper must clarify which quantity is actually implemented. This matters because the abstract's third contribution is that the divergence-free condition is 'built into the optimization' so that spurious eigenpairs are 'automatically filtered'. If the actual ratio is the seminorm ratio, the separation argument for spurious modes changes: a curl-free spurious mode can have small divergence, so no uniform lower bound on the ratio follows. If the actual ratio is the full-norm ratio, then real low eigenvalues incur a nontrivial penalty and the paper's explanation is wrong. Either way, the central filtering claim is not established by the text. The numerical results themselves are strong and demonstrate accuracy on several benchmarks, and a corrected norm statement or a threshold analysis would likely preserve the method's practical value, which is why this does not change the reader's conditional verdict.","tokens_in":39,"tokens_out":6906,"duration_ms":131731,"concrete_test":"Recompute the filter ratio for the first square-cavity true eigenpair E=(0, sinπx1): under §2, ‖E‖²_H(div)=1/2 and ‖E‖²_H(curl)=(1+π²)/2, so ρ=1/(1+π²), not near zero. Then rerun the square-cavity experiment with loss (5.13) implemented once with the full norms of §2 and once with the seminorm variant. If the code actually uses the seminorm (as Table 6.1's ~1e-9 H(div) values indicate), report that explicitly and re-derive the spurious-mode separation; if it uses full norms, show how the threshold separates ρ≈0.092 from ρ≥1 in practice. Additionally, insert a synthetic two-dimensional curl-free, non-divergence-free eigenspace into the generalized problem (5.4) and check whether both spurious eigenvectors are rejected by the loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The divergence-free filtering mechanism in §5.2.1 is not justified as written. With the norms defined in §2, ρ(E)=‖E‖²_H(div)/‖E‖²_H(curl) = (‖E‖²_L2+‖divE‖²)/(‖E‖²_L2+‖curlE‖²). For a normalized real eigenpair, A(E,E)=‖curlE‖²=λ‖E‖², so ρ=1/(1+λ), not \"close to 0\"; for the first square-cavity mode λ=π², ρ≈0.092. For a curl-free spurious mode, the denominator reduces to ‖E‖², so ρ≥1. Thus the ratio separates a low-λ real mode from a spurious mode only by a factor λ+1, and the claim that the numerator \"approaches 0\" is incompatible with the stated H(div) norm, which contains the L2 term. The tables report |E_NN|_H(div) ~1e-9 for normalized eigenfunctions, which is impossible under that same norm; the text must mean the seminorm ‖divE‖. If the seminorm is intended, separation for spurious modes is no longer guaranteed merely by \"curl-free\", since a curl-free field can have arbitrarily small divergence. Moreover, the cited curl-free property is only stated for simple spurious eigenvalues (§2, §5.2.1); no analysis or experiment covers multiple or near-degenerate spurious modes. Since the central claim is \"automatic filtering of spurious eigenpairs\", this is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FieldTNN, a tensor neural network architecture for vector-valued functions, and applies it to the Maxwell eigenvalue problem in two and three dimensions on both tensor-product and non-tensor domains. For non-tensor domains, the method uses locally compactly supported FieldTNN bases on subdomains. The authors propose a loss function that adds a penalty proportional to a H(div)/H(curl) ratio of the computed eigenfunctions in order to automatically filter spurious, non-divergence-free eigenpairs. Numerical experiments on square, L-shaped, inhomogeneous, cubic, and 3D L-shaped cavities report very high accuracy for tensor domains and moderate accuracy for singular non-tensor domains, with comparisons to existing finite element and spectral element results.","tokens_in":23334,"tokens_out":2798,"duration_ms":26759,"significance":"If the claims are substantiated, this is a useful extension of the TNN framework to vector field eigenvalue problems and to non-tensor computational domains. The reported tensor-domain accuracy is genuinely strong: relative eigenvalue errors around 1e-8 to 1e-9 in the square and cube cavities, competitive with or better than the cited standard methods, and the quadrature splitting in Section 5.1 provides a concrete complexity argument for avoiding tensor-product quadrature in low-rank formats. The treatment of non-tensor domains through locally supported TNN bases is a reasonable and potentially transferable idea. However, the central claim that spurious eigenpairs are automatically filtered by the proposed ratio-based penalty is not established as written; the mathematical justification in Section 5.2.1 is inconsistent with the norms defined in Section 2 and does not cover the reported numerical evidence.","major_comments":[{"comment":"The filter ratio as written is incompatible with the norms defined in Section 2. With the full H(div) and H(curl) norms, a normalized real eigenpair satisfies rho = ||E||^2_H(div)/||E||^2_H(curl) = (1 + ||div E||^2)/(1 + ||curl E||^2) = 1/(1 + lambda), which is not close to zero for small eigenvalues; for the first square-cavity mode lambda = pi^2 this gives approximately 0.092. Yet the tables report |E_NN|_H(div) around 1e-9 for normalized eigenfunctions, which is impossible under that same norm because the L2 component alone is of order 1. The text must mean the seminorm ||div E||_L2, not the H(div) norm. This needs to be stated explicitly and consistently throughout, and the separation mechanism must be re-derived for the seminorm, since the claim that the numerator approaches zero is otherwise false.","section":"Section 5.2.1, Eq. (5.13), Tables 6.1, 6.3, 6.5, 6.7, 6.9"},{"comment":"The justification for separating spurious from real eigenpairs is incomplete. The cited curl-free property is stated only for simple spurious eigenvalues, and the argument that the denominator of the ratio approaches zero relies on that property. For multiple or near-degenerate spurious eigenvalues, no analysis is provided, and a curl-free field can have arbitrarily small divergence, so the ratio with the seminorm does not guarantee a large value for all spurious modes. Since the paper's advertised mechanism is 'automatic filtering of spurious eigenpairs', the authors should either prove the separation under the intended seminorm, or provide a numerical study covering clustered and multiple spurious eigenvalues.","section":"Section 5.2.1, Section 2"},{"comment":"The penalty parameter beta in the loss function (5.13) is never disclosed, and the text does not explain how beta is chosen or how the penalty interacts with the generalized eigenvalue problem solved in post-processing. The final eigenvalues are extracted from the GEVP (5.4), not directly from the minimized loss, so it is unclear whether the penalty term actually changes the eigenfunction basis during training or merely adds a constant that vanishes for divergence-free fields. The authors should report beta for each numerical example, discuss its influence on the computed eigenvalues and on the reported H(div) values, and ideally give an error or consistency analysis showing that the penalty does not bias the converged eigenpairs.","section":"Sections 5.3 and 6, Eq. (5.13)"}],"minor_comments":[{"comment":"The name is misspelled as 'FiedTNN' in the text introducing the numerical results; it should be 'FieldTNN'.","section":"Sections 6.1, 6.3"},{"comment":"Algorithm 2 has two steps labeled 'Step 1'; the second one should be renumbered.","section":"Algorithm 2, Section 5.3"},{"comment":"The word 'filter' is misspelled as 'fileter' in 'fileter ratio'.","section":"Section 5.2.1"},{"comment":"The caption contains the typo 'pruple', which should be 'purple'.","section":"Figure 3.2 caption"},{"comment":"The numerical setup for the 2D L-shaped cavity reports only Adam optimization, whereas Sections 6.1 and 6.4 describe an additional L-BFGS refinement phase; it would be helpful to state whether such refinement was used or omitted for consistency.","section":"Section 6.2"}],"recommendation":"major_revision","confidential_remarks":"The core numerical results in tensor domains are impressive and likely publishable after the filtering mechanism is clarified. The main risk is that the divergence-free filtering claim, which is the headline contribution regarding spurious modes, is not supported by the mathematics as presented. This is fixable in revision by switching to the seminorm, disclosing beta, and adding either a proof or a targeted experiment for the separation property, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The FieldTNN extension to vector fields is a real step forward: tensor-domain results (square, cube) are impressively accurate, with relative errors around 1e-8 to 1e-9 and correct handling of multiplicities. The non-tensor domain construction via compactly supported subnetworks is also a genuinely useful idea, and the L-shaped and inhomogeneous examples show the method works beyond tensor boxes, even if errors there are more modest (1e-3 for the singular first 2D L-shaped mode). Comparisons against two-grid edge elements and mixed spectral elements are fair and the method comes out competitive or better.\n\nThe soft spots are real. Section 5.2.1 defines the filter ratio using the full H(div) and H(curl) norms from Section 2, which include the L2 term. Under those norms, a real eigenpair with lambda = pi^2 has rho = 1/(1+lambda) ≈ 0.09, not 'close to 0,' and a curl-free spurious mode has rho ≥ 1, not 'large' because the denominator vanishes. The tables report |E_NN|_H(div) values around 1e-10, which is impossible if the norm includes the L2 part. The authors must be using the seminorm, but they never say so. This is not a typo: the entire automatic-filtering argument rests on this ratio. If the seminorm is intended, the separation is no longer automatic — a curl-free field can have small divergence, and no analysis or experiment addresses multiple or near-degenerate spurious modes. The penalty parameter beta is also never disclosed, which hurts reproducibility.\n\nThat said, the numerical evidence suggests the method does filter spurious modes in practice, and the generalized eigenproblem (5.4) at the end is a sensible extraction mechanism. I do not think the paper's results are wrong, but the explanation of the mechanism is inconsistent with its own definitions and needs fixing. Lack of error analysis and code is a lesser concern for a methods paper, but worth noting.\n\nThis paper is for researchers working on ML-based PDE solvers, especially vector field problems or non-tensor domains. It deserves a serious referee: the idea is novel and the numerics are strong, but the referee should push for a clean statement of the filter ratio and at least a numerical study of degenerate spurious modes. Send it to review.","headline":"Impressive numerics for Maxwell eigenvalues with FieldTNN, but the divergence-free filter's justification doesn't match the paper's own norm definitions.","tokens_in":23718,"tokens_out":3856,"would_cite":true,"duration_ms":35690,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35A15","35Q60","65N25","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"FieldTNN solves Maxwell eigenvalue problems in tensor and non-tensor cavities, with spurious modes filtered automatically by a divergence-free penalty in the loss.","keywords":["Maxwell eigenvalue problem","FieldTNN","tensor neural network","divergence-free condition","spurious eigenpairs","non-tensor domain","machine learning for PDEs","electromagnetic cavity"],"falsifier":"A concrete test: construct a Maxwell eigenvalue problem whose relaxed variational form (2.2) has a multiple spurious eigenvalue, or seed the FieldTNN ansatz with two linearly independent curl-free vector fields of equal Rayleigh quotient, and train with loss (5.13). If any reported eigenpair has a large curl-free component yet still attains a small loss and a small ratio $\\rho$, the filter has misclassified it; the paper provides no theoretical or numerical evidence for this degenerate case.","tokens_in":22649,"feed_emoji":"⚡","tokens_out":8553,"duration_ms":69304,"temperature":0.7,"pith_summary":"The paper proposes a tensor-neural-network method, called FieldTNN, for computing the leading eigenvalues and eigenfunctions of the Maxwell eigenvalue problem in 2D and 3D electromagnetic cavities, including L-shaped and inhomogeneous domains. Its central promise is that the divergence-free constraint—the main source of spurious modes in discretizations of Maxwell's equations—can be built directly into the loss function used for training, so that unphysical eigenpairs are filtered out automatically instead of being removed by a separate post-processing step. The method also extends tensor neural networks from rectangular tensor-product domains to non-tensor domains by decomposing the domain into subdomains and training compactly supported basis functions on each piece. The paper reports relative eigenvalue errors below $10^{-8}$ on smooth cavities and demonstrates usable accuracy on singular L-shaped domains, with the divergence norm of computed eigenfunctions small in all examples. If the approach holds up, it offers a mesh-free alternative for cavity-resonance computations in complex geometries.","feed_headline":"FieldTNN computes cavity resonances without spurious modes","feed_subtitle":"A tensor neural network embeds a divergence-free penalty in its loss, so only genuine Maxwell eigenpairs survive.","key_machinery":"The central object is the FieldTNN itself: a rank-$p$ vector function $\\Psi(x;\\Theta) = \\sum_{k=1}^p u_k \\hat\\Psi_k(x;\\Theta)$, where each vector-valued rank-one block $\\hat\\Psi_k$ has components that are products of one-dimensional subnetworks, one subnetwork per coordinate per field component. This product structure lets the curl and divergence of the network be written as sums of products of one-dimensional functions, so the stiffness and mass matrices $S_{ji}=A(\\hat\\Phi_i,\\hat\\Phi_j)$ and $M_{ji}=B(\\hat\\Phi_i,\\hat\\Phi_j)$ of the Galerkin discretization factor into one-dimensional integrals and are assembled by high-order one-dimensional quadrature in $O(d^3 M N)$ operations. The machinery that carries the spurious-mode filtering argument is the ratio $\\rho = |E_{NN}|^2_{H(\\mathrm{div};\\Omega)}/|E_{NN}|^2_{H(\\mathrm{curl};\\Omega)}$ added as a penalty in the loss (5.13). For non-tensor domains, the paper constructs compactly supported subnetworks by composing each subnetwork with the one-dimensional clamping function $g_{[a,b]}(x)$ built from ReLU units, so that basis functions vanish outside their subdomain and global matrices are assembled by summing subdomain contributions. The boundary condition $E\\times n=0$ is enforced by multiplying each field component by analytic factors $\\gamma_i(x_i)$ that vanish on the appropriate faces.","core_discovery":"The paper's central claim is that a FieldTNN—a vector-valued tensor neural network in which each component of the field is a sum of products of one-dimensional subnetworks—can approximate Maxwell eigenpairs to high accuracy while automatically excluding spurious ones. The key mechanism is a loss function $L_{\\mathrm{Maxwell}}(\\lambda_{k,NN}, E_{k,NN}) = \\lambda_{k,NN} + \\beta |E_{k,NN}|^2_{H(\\mathrm{div};\\Omega)} / |E_{k,NN}|^2_{H(\\mathrm{curl};\\Omega)}$ applied to the leading generalized eigenpairs of the FieldTNN Galerkin system. Because genuine eigenfunctions are divergence-free, their $H(\\mathrm{div})$ norm is small, while spurious eigenpairs of the relaxed variational formulation are curl-free when simple, making their $H(\\mathrm{curl})$ norm small and the ratio large. Minimizing the summed loss therefore drives the network toward divergence-free eigenspaces, and the paper's numerical experiments show no spurious eigenvalues surviving. On tensor domains the basis functions are products of subnetworks over coordinate intervals; on non-tensor domains the domain is split into tensor subdomains and locally supported subnetworks are trained in each, mimicking spectral element assembly. The paper claims this is the first application of tensor neural networks to vector field eigenvalue problems and the first extension of the TNN framework to non-tensor computational domains.","pith_inferences":["Editorial inference: The divergence-free filter relies on the sharp distinction between divergence-free real modes and curl-free simple spurious modes; if a spurious eigenvalue is multiple or nearly degenerate, the ratio $\\rho$ may not separate eigenpairs, and the method would need a stabilizer or deflation to stay reliable.","Editorial inference: The penalty parameter $\\beta$ creates a trade-off: too small leaves spurious modes with small loss, too large distorts the eigenvalue estimate; a principled schedule for $\\beta$ is a natural next test.","Editorial inference: One could test the method on a domain with a reentrant corner of different angle, where the singular exponent of the Maxwell eigenfunction changes, to see whether the reported accuracy on the 90-degree L-shape degrades.","Editorial inference: Because the boundary condition is enforced by analytic multipliers rather than by the network itself, the method's extension to curved or material-interface boundaries would require rederiving those multipliers, which is a concrete limitation not addressed in the paper."],"forward_implications":["Leading Maxwell eigenvalues in smooth 2D and 3D cavities can be computed with relative errors around $10^{-9}$ to $10^{-8}$, competitive with or better than the two-grid edge-element and mixed spectral element comparisons in the paper.","Spurious eigenpairs are filtered during training rather than in a separate post-processing pass, so the output eigenpairs satisfy the divergence-free condition to the tolerance shown by small $|E_{NN}|_{H(\\mathrm{div};\\Omega)}$ values.","The method applies to non-tensor domains such as 2D and 3D L-shaped cavities, where singular eigenfunctions with unbounded gradients are still captured, with largest relative eigenvalue errors around $10^{-3}$ for the singular leading mode.","The same FieldTNN construction is claimed to extend to other vector field equations—Maxwell equations, Stokes equations, magnetohydrodynamics—since the architecture and quadrature splitting depend only on the product structure of the field components.","Because integration is reduced to one-dimensional quadrature with polynomial complexity in dimension, the approach avoids the curse of dimensionality for the variational forms used here."],"supporting_citations":[{"why":"Supplies the property that simple spurious eigenpairs of the relaxed Maxwell formulation are curl-free, which underpins the filter ratio.","marker":"[5]"},{"why":"Establishes the occurrence of spurious eigenpairs in discretizations of (2.2) and the three computational strategies the paper adapts.","marker":"[13]"},{"why":"Introduces the tensor neural network architecture and the numerical integration scheme that the FieldTNN quadrature extends.","marker":"[27]"},{"why":"Provides the TNN method for computing multiple eigenpairs of high-dimensional eigenvalue problems, which the paper generalizes to vector fields.","marker":"[29]"},{"why":"Inspires the domain-decomposition strategy with locally supported basis functions that the paper uses for non-tensor domains.","marker":"[14]"},{"why":"Supplies the two-grid edge-element results used as a comparison baseline for square and cube cavities.","marker":"[30]"},{"why":"Supplies the mixed spectral element results used as comparison baselines for L-shaped and inhomogeneous cavities.","marker":"[20]"},{"why":"Provides the one-dimensional quadrature theory used for the factored integrations in the stiffness and mass matrices.","marker":"[25]"}],"fun_headline_variants":["FieldTNN solves Maxwell eigenproblems, filters spurious modes","Tensor NN method for Maxwell eigenproblems with no spurious modes","FieldTNN extends TNN to non-tensor domains for Maxwell eigenproblems","Divergence penalty in FieldTNN kills spurious Maxwell eigenpairs","FieldTNN handles non-tensor domains for Maxwell eigenvalue problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's ability to throw out fake solutions depends on a clean split: true eigen-solutions carry no divergence, while simple fake solutions carry no curl, so the ratio in the loss separates them; if a fake eigenvalue is repeated or mixed with a real one, that split is no longer guaranteed and the paper gives no analysis for that case.","fun_headline_variants_meta":{"raw":{"variants":["FieldTNN solves Maxwell eigenproblems, filters spurious modes","Tensor NN method for Maxwell eigenproblems with no spurious modes","FieldTNN extends TNN to non-tensor domains for Maxwell eigenproblems","Divergence penalty in FieldTNN kills spurious Maxwell eigenpairs","FieldTNN handles non-tensor domains for Maxwell eigenvalue problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3215,"prompt_tokens":957,"completion_tokens":2258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2166}},"tokens_in":573,"tokens_out":2258,"duration_ms":13997,"temperature":1.0,"reasoning_tokens":2166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:50:44.880658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: construct a Maxwell eigenvalue problem whose relaxed variational form (2.2) has a multiple spurious eigenvalue, or seed the FieldTNN ansatz with two linearly independent curl-free vector fields of equal Rayleigh quotient, and train with loss (5.13). If any reported eigenpair has a large curl-free component yet still attains a small loss and a small ratio $\\rho$, the filter has misclassified it; the paper provides no theoretical or numerical evidence for this degenerate case.","supporting_citations":[{"cited_title":"Costabel and M","cited_arxiv_id":null,"evidence_quote":"Supplies the property that simple spurious eigenpairs of the relaxed Maxwell formulation are curl-free, which underpins the filter ratio."},{"cited_title":"Ciarlet Jr and G","cited_arxiv_id":null,"evidence_quote":"Establishes the occurrence of spurious eigenpairs in discretizations of (2.2) and the three computational strategies the paper adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the tensor neural network architecture and the numerical integration scheme that the FieldTNN quadrature extends."},{"cited_title":"Wang and H","cited_arxiv_id":null,"evidence_quote":"Provides the TNN method for computing multiple eigenpairs of high-dimensional eigenvalue problems, which the paper generalizes to vector fields."},{"cited_title":"Kwan and J","cited_arxiv_id":null,"evidence_quote":"Inspires the domain-decomposition strategy with locally supported basis functions that the paper uses for non-tensor domains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the two-grid edge-element results used as a comparison baseline for square and cube cavities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mixed spectral element results used as comparison baselines for L-shaped and inhomogeneous cavities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the one-dimensional quadrature theory used for the factored integrations in the stiffness and mass matrices."}],"review_version":1}