{"id":"029fd02f-75ef-405d-bfc5-4fbc8358be1e","arxiv_id":"2507.05786","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural network that approximates virtual element basis functions from boundary data yields a stabilization-free polygonal discretization that converges on linear and nonlinear elasticity benchmarks.","lead":"This paper extends a neural-network variant of the Virtual Element Method to linear and nonlinear elasticity, replacing the usual stabilization and projection operators with element-wise neural approximations of the basis functions. The approach is tested on convergence benchmarks, a nonlinear material with manufactured solutions, and a Neo-Hookean hyperelastic problem, reporting accuracy competitive with or better than standard VEM.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence evidence is in-distribution only: all test meshes are admitted to be well represented by the training set, leaving the central claim of a reusable, general polygonal method unsupported.","rationale":"The reader's weakest assumption combines the boundary-to-interior transfer theorem from [22] with the empirical generalization of the trained networks. I agree that the transfer theorem is a necessary condition, but it is at least supported by a proof in the authors' earlier work and by the small boundary losses reported in Table 1. The more actionable and less supported condition is generalization: the paper itself concedes in Section 4.1 that the test meshes are 'well represented by the training set,' which means the convergence experiments do not exercise the network outside its training distribution. This is a testable empirical gap rather than a demonstrated inconsistency, so the appropriate verdict remains CONDITIONAL: the central claim should not be accepted as a general method until out-of-distribution meshes are tested. I would not reject the paper, because the method may well generalize; the evidence is simply incomplete. The concrete verification I propose would settle the issue by measuring boundary losses and convergence on deliberately dissimilar mesh families.","tokens_in":17341,"tokens_out":6563,"duration_ms":77579,"concrete_test":"Evaluate the pretrained networks for Nv=4 and Nv=6 on a mesh family that is not drawn from RDQM or VM: for example, (i) quadrilaterals with aspect ratio 10:1 or with one very small angle (below 10 degrees), (ii) non-convex hexagons, and (iii) meshes with mixed vertex counts. For each family, compute the actual boundary losses (18) and (22) on the test elements and the H1 error (29) for Test 1's linear problem with the same Lame parameters. If the boundary losses remain comparable to Table 1 and O(h) convergence is observed, the generalization concern is resolved; if the losses spike or convergence degrades, the method's central claim is restricted to the training distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NA VEM provides a general, stabilization-free polygonal discretization whose pre-trained networks can be reused on various meshes (Introduction, Section 3.3). The numerical support in Section 4 consists of convergence tests on two mesh families: randomly distorted quadrilateral grids and distorted Voronoi meshes. Section 4.1 explicitly states that 'the elements in this family of meshes are well represented by the training set. Thus, the network accuracy does not influence these results.' This is an admission that the reported convergence rates are in-distribution performance. The training datasets RDQM and VM are the same distributions from which the test meshes are drawn, so the experiments cannot distinguish between a generalizable basis-function approximator and a network that has effectively memorized the training distribution. If the latter is true, the method would not deliver the claimed offline-online advantage on genuinely new meshes, and the stabilization-free convergence would not be assured. This concern is load-bearing because the entire premise of the method is that the same networks can be applied to arbitrary polygons of a given vertex count, not merely to polygons statistically identical to the training set.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces the Neural Approximated Virtual Element Method (NA VEM) for linear and nonlinear elasticity in two dimensions. The key idea is to replace the implicitly defined VEM basis functions with explicit neural-network approximations that are trained once, in an offline phase, to match the known boundary traces of the lowest-order VEM basis functions on reference polygons. In the online phase the networks return closed-form basis functions and their gradients, so the discrete equations are assembled as in a standard finite element method, without stabilization or projection operators. The paper presents the discrete formulation, details the network architecture and training losses, and reports numerical tests for linear elasticity, a nonlinear benchmark with analytical solutions, and a Neo-Hookean hyperelasticity problem. The central claims are that NA VEM attains the expected O(h^2) L2 and O(h) H1 convergence rates and achieves lower error constants than VEM, while avoiding stabilization and being simpler in nonlinear regimes.","tokens_in":17594,"tokens_out":5414,"duration_ms":65096,"significance":"If the claims are substantiated, NA VEM would be a practically attractive stabilization-free polygonal method for nonlinear elasticity, with the possible offline-online advantage that the same trained networks can be reused across different meshes, materials, and constitutive laws. The paper has several genuine strengths: the benchmarks use external analytical solutions, the nonlinear tests are substantially more demanding than a single linear Poisson problem, comparisons are made against independent VEM and FEM results, and no free parameter is fit to the PDE solution in the training procedure, so the experiments are not circular. The improved training losses and optimizer are also documented quantitatively in Table 1. However, the two load-bearing numerical claims—lower error constants than VEM and reusability on general meshes—are not fully supported by the evidence as presented: the error metrics used for NA VEM and VEM are different, and the test meshes are drawn from the same distributions used for training. These issues are fixable but require additional experiments and clarification.","major_comments":[{"comment":"The claimed lower error constants of NA VEM relative to VEM are not established because the two methods are measured with different error functionals. Equation (29) compares the NA VEM solution directly pointwise (u - u_NN), while Eq. (30) compares VEM through the projected quantities u - Π1 u_VEM and ∇u - Π0 ∇u_VEM. Projection is a smoothing operation, so a lower value of the VEM error functional does not necessarily correspond to a more accurate VEM solution; conversely, the comparison between the two functionals is not apples-to-apples. To support the 'lower error constants' claim, the authors should compare both methods on the same quantity, for example by applying the same projection operators to the NA VEM solution, or by using an enhanced VEM solution that can be evaluated pointwise. The paper should also report computed convergence rates (EOC values) rather than relying only on visual inspection of log-log plots.","section":"Section 4, Eqs. (29)-(30)"},{"comment":"The numerical evidence for the central claim of a reusable, general polygonal method is in-distribution only. The training datasets RDQM and VM used in Section 3.3 are the same distributions from which the test meshes in Section 4.1 and 4.2 are drawn, and the manuscript explicitly states that 'the elements in this family of meshes are well represented by the training set. Thus, the network accuracy does not influence these results.' This is an admission that the reported convergence rates do not test generalization to genuinely new polygon shapes. Since the introduction and Section 3.3 present the offline-online reuse across 'various meshes' as a key advantage, the authors should include out-of-distribution experiments, for example polygonal meshes with strong aspect-ratio distortion, non-convex polygons, or vertex counts outside the trained classes, and report the resulting errors. Without such experiments, the claim that the same networks can be applied to arbitrary polygons of a given vertex count is unsupported.","section":"Section 4.1, paragraph beginning 'We also note that the elements...'"},{"comment":"The NA VEM discrete space is nonconforming: the manuscript acknowledges that the neural-network basis functions are not exactly continuous across elements, yet the degrees of freedom are not doubled. No consistency estimate or patch test is provided for this nonconforming approximation. Since the method is presented as a finite-element-like discretization, the observed convergence needs to be supported by either a standard patch test or a discussion of the consistency error introduced by the trace mismatch. As it stands, the convergence rates in Section 4 could in principle be masking a small but uncontrolled variational crime, especially on meshes that are less well represented by the training set.","section":"Section 3.4, Eqs. (24)-(25)"},{"comment":"The convergence-rate claims O(h^2) and O(h) are supported only by visual inspection of log-log plots; no computed experimental orders of convergence are tabulated. Moreover, in Test 2 the L2 errors are said to behave similarly but are not reported. The authors should provide a table or listing of computed EOC values for each mesh family, error norm, and test case. This is essential for verifying that the convergence is actually of the expected order and for comparing the two methods fairly.","section":"Section 4, Figures 2, 4, 5"}],"minor_comments":[{"comment":"In Eq. (13), the phrase 'with = 1, . . . , Nv_E' is missing the index variable; it should read 'with j = 1, . . . , Nv_E'.","section":"Section 3.1, Eq. (13)"},{"comment":"The text says 'his gradient can be approximated'; this should be 'this gradient can be approximated'.","section":"Section 3.2, paragraph after Eq. (18)"},{"comment":"The caption of Figure 5 says 'Figure 4 reports the behaviour...' but should refer to Figure 5 itself.","section":"Section 4.2, caption of Figure 5"},{"comment":"The input encoding strategy and the affine map from a general polygon to the reference square are described only by reference to [22]. Since the encoding is essential for reproducing the method, a brief self-contained description or an appendix entry would improve the paper.","section":"Section 3.2 and 3.3"},{"comment":"The mesh sizes for the mesh in Figure 7 and the FEM reference mesh in Figure 8 are not stated, which makes the visual comparison of oscillations difficult to interpret quantitatively.","section":"Section 4.3, Figures 7 and 8"}],"recommendation":"major_revision","confidential_remarks":"This is a promising manuscript and the experiments are more demanding than many papers in this area. The main risks are not circularity but (i) the unfair error comparison between NA VEM and VEM and (ii) the absence of out-of-distribution generalization tests. Both are fixable within the scope of a revision. I would encourage the editor to ask for the additional experiments and a tabulation of convergence rates before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a competent extension of the authors' NA VEM from scalar elliptic problems to linear and nonlinear elasticity. The vector-valued construction, the treatment of hyperelastic materials without projectors, and the improved training procedure (loss and BFGS phase) are genuinely new relative to their earlier papers. The numerical results show the expected O(h^2) and O(h) behavior on the tested meshes, and the large-deformation neo-Hookean example is a meaningful stress test.\n\nThe method is sensible: approximating virtual basis functions by networks trained on boundary traces, then using them pointwise, removes the stabilization term, which is the most ad hoc part of VEM for nonlinear problems. The authors are also honest about the key caveat: they state in Section 4.1 that the test meshes are well represented by the training set, so the reported convergence is in-distribution performance. That is not a fatal flaw—the training data is geometric, not problem-dependent, and the benchmarks are external analytical solutions—but it does mean the headline claim of reusability on arbitrary polygons is not actually demonstrated.\n\nThree soft spots are worth naming, in order of importance. First, the comparison of error constants between NA VEM and VEM is not apples-to-apples: NA VEM errors are direct L2 and H1 norms of the full solution, while VEM errors use projected quantities (equations 29 vs 30). This can make NA VEM look better relative to VEM than it actually is. Second, no computed convergence rates are tabulated; only plots are shown. Third, no code or data is released, so the results are not independently reproducible. The stress-test note about out-of-distribution generalization is correct—highly anisotropic or badly shaped polygons are not tested—but the authors do flag the in-distribution nature of their meshes, which counts in their favor.\n\nWho is this for? Researchers in computational mechanics and SciML interested in stabilization-free polygonal methods. The paper is a reasonable contribution, not a breakthrough. It deserves serious peer review, with requests for code/data, a direct error-metric comparison, and at least one out-of-distribution generalization test. I would read a revised version carefully.","headline":"Solid extension of NA VEM to elasticity with honest but in-distribution numerical evidence; the generalization claim is plausible but unproven, and the error-constant comparison is muddied by different error metrics.","tokens_in":18089,"tokens_out":1979,"would_cite":true,"duration_ms":23455,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N12","74S05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural-network-predicted basis functions replace stabilization in a polygonal method for elasticity, matching expected convergence orders.","keywords":["NA VEM","virtual element method","neural network approximation","elasticity","polygonal meshes","stabilization-free","nonlinear elasticity","basis function approximation"],"falsifier":"Evaluate the trained networks on polygons with extreme aspect ratios or non-convex shapes outside the RDQM and VM training distributions; the central claim is falsified if the $L^2$ or $H^1$ errors stop decreasing at the expected rates or the convergence curves level off.","tokens_in":1868,"feed_emoji":"🧠","tokens_out":4263,"duration_ms":81136,"temperature":0.7,"pith_summary":"This paper extends the Neural Approximated Virtual Element Method (NA VEM) to linear and nonlinear elasticity. It claims that virtual basis functions can be approximated element-wise by neural networks trained only on boundary data, so that no stabilization or projection operators are needed. On distorted quadrilateral and Voronoi meshes, the method reaches the expected convergence orders, $O(h^2)$ in the $L^2$ error and $O(h)$ in the $H^1$ seminorm, with lower error constants than the standard virtual element method. On a large-displacement nonlinear benchmark, it stays stable and oscillation-free even when the standard method requires an incremental loading strategy to maintain accuracy.","feed_headline":"Neural basis functions reach VEM accuracy without stabilization","feed_subtitle":"NA VEM predicts virtual basis functions element-wise and matches expected convergence rates on elasticity tests.","key_machinery":"The local approximation space is $H^{\\mathrm{NN}}_{j,E} = \\mathrm{span}\\{\\tilde p_\\beta\\}_{\\beta=1}^{2\\ell_{\\mathrm{NN}}+1} \\cup \\{\\Phi^{j-1}_{j,E}, \\Phi^j_{j,E}, \\Phi^{j+1}_{j,E}\\}$, where the polynomials are harmonic polynomials up to degree $\\ell_{\\mathrm{NN}}$ and the three auxiliary functions are built from a universal harmonic-rational function $\\Phi(z)$ with poles $z_\\alpha = 1 + 2\\exp(-4(\\sqrt{N_1}-\\sqrt{\\alpha}))$. Two neural networks map an encoding of the pair (vertex, element) to coefficients: $c^\\varphi$ for the basis function via loss (18) and $c^q$ for the gradient via loss (22). The load-bearing transfer principle, proven in [22], is that small boundary losses suffice for accurate interior approximation of both the function and its gradient. Training uses an RMS loss and a two-stage optimization with ADAM followed by a self-scaled BFGS optimizer, which the paper reports lowers the training losses compared with [22]. The gradient network is deliberately not the derivative of the basis-function network; this introduces a consistency error but suppresses gradient oscillations and makes the method more stable.","core_discovery":"The authors construct two small neural networks per polygon-vertex class: one predicts the coefficients of a basis-function approximation in a space spanned by harmonic polynomials plus three element-dependent auxiliary harmonic functions, and a second predicts coefficients for the gradient of that basis function. Both networks are trained by minimizing boundary trace losses, namely the $H^{1/2}(\\partial E)$ loss (18) for the function and the tangential $L^2(\\partial E)$ loss (22) for the gradient, using the fact that virtual basis functions are known in closed form on element boundaries. The paper relies on the result from [22] that small boundary losses guarantee accurate interior approximation of both the basis functions and their gradients. The assembled discrete problem then follows the standard finite element structure, with no projector or stabilization term. Numerical tests on linear elasticity, on the nonlinear benchmark from [10], and on a Neo-Hookean hyperelastic problem show the expected convergence rates and lower error constants relative to VEM, and the large-deformation test converges without the incremental force strategy that VEM requires.","pith_inferences":["Editorial extension: if the boundary-to-interior transfer holds beyond the tested mesh families, the method could be applied to non-convex or badly shaped elements, provided the training set adequately covers such geometries.","Editorial extension: the generalization of the trained networks to polygons outside the RDQM and VM training distributions is untested; a natural stress test is to evaluate on extreme-aspect-ratio or non-convex elements, where boundary trace losses may not control interior gradients.","Editorial extension: the independence of the basis-function and gradient networks creates an inconsistency that may influence stress accuracy; the paper reports stability benefits but does not analyze the resulting consistency error.","Editorial extension: because the approximation is element-wise and vertex-local, the method could combine naturally with adaptive refinement or hanging-node strategies, as the authors note in their conclusions."],"forward_implications":["The same trained networks can be reused for different materials, constitutive laws, body forces, and meshes whose polygons have the same vertex count, because training uses only geometric information.","Nonlinear problems become simpler: no stabilization parameter needs tuning, and the large-displacement benchmark converges with plain Newton-Raphson, without the incremental force strategy that VEM needs.","Point-wise evaluation of basis functions and gradients is available everywhere, so quantities such as the deformation determinant can be computed directly rather than through a projector.","The method retains the expected finite element convergence rates on polygonal meshes, with lower error constants than VEM in the reported tests.","The per-element neural network evaluation cost becomes negligible on fine meshes, and is further amortized in vector-valued elasticity because scalar basis functions are reused across components."],"supporting_citations":[{"why":"Establishes the NA VEM framework and proves that small boundary losses ensure accurate interior approximation of basis functions and gradients.","marker":"[22]"},{"why":"Defines the virtual element space for linear elasticity whose basis functions NA VEM approximates.","marker":"[9]"},{"why":"Provides the VEM formulation for elastic and inelastic problems, the stabilization choices, and the nonlinear benchmark tests used as comparison.","marker":"[10]"},{"why":"Introduces the scalar virtual element space and its Lagrange basis on polygons, which underpins the local approximation target.","marker":"[7]"},{"why":"Discusses VEM for finite deformations and determinant approximation, which NA VEM simplifies by point-wise evaluation.","marker":"[14]"},{"why":"Supplies the Voronoi mesh generator used both for the VM training datasets and for the distorted Voronoi test meshes.","marker":"[38]"},{"why":"Provides the self-scaled BFGS optimizer whose second training stage reduces the boundary losses reported in Table 1.","marker":"[37]"},{"why":"Earlier NA VEM formulation for elliptic problems, extended here to elasticity, and source of the boundary-loss training strategy.","marker":"[21]"}],"fun_headline_variants":["Neural virtual elements skip stabilization in elasticity","Neural networks replace VEM stabilization for elasticity","Neural VEM hits expected accuracy without projectors","Neural virtual elements cut stabilization, handle nonlinearity"],"cache_read_input_tokens":20224,"weakest_assumption_plain":"The load-bearing premise is that minimizing the boundary trace losses (18) and (22) is enough to guarantee accurate interior approximation of both the basis functions and their gradients on any polygon of a given vertex count, and that networks trained on the RDQM and VM datasets generalize to unseen polygons of the same vertex count.","fun_headline_variants_meta":{"raw":{"variants":["Neural virtual elements skip stabilization in elasticity","Neural networks replace VEM stabilization for elasticity","Neural VEM hits expected accuracy without projectors","Neural virtual elements cut stabilization, handle nonlinearity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000542,"raw_usage":{"total_tokens":2541,"prompt_tokens":836,"completion_tokens":1705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1645}},"tokens_in":452,"tokens_out":1705,"duration_ms":16529,"temperature":1.0,"reasoning_tokens":1645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:18:12.977912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained networks on polygons with extreme aspect ratios or non-convex shapes outside the RDQM and VM training distributions; the central claim is falsified if the $L^2$ or $H^1$ errors stop decreasing at the expected rates or the convergence curves level off.","supporting_citations":[{"cited_title":"Berrone, M","cited_arxiv_id":null,"evidence_quote":"Establishes the NA VEM framework and proves that small boundary losses ensure accurate interior approximation of basis functions and gradients."},{"cited_title":"Beirão da Veiga, F","cited_arxiv_id":null,"evidence_quote":"Introduces the scalar virtual element space and its Lagrange basis on polygons, which underpins the local approximation target."},{"cited_title":"Berrone, D","cited_arxiv_id":null,"evidence_quote":"Earlier NA VEM formulation for elliptic problems, extended here to elasticity, and source of the boundary-loss training strategy."}],"review_version":1}