{"id":"05ec951d-0a29-4f89-ba17-bb72a9aca749","arxiv_id":"2501.06925","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural network trained on 80 Virtual Element Method solutions is offered as a beam displacement surrogate, but its claimed accuracy, generalization, and speed are unverified against external references.","lead":"This paper trains a neural network on Virtual Element Method solutions to predict bending displacements for Euler-Bernoulli beams and frames. It combines two existing training techniques, Sobolev training and GradNorm, but reports no code, baselines, or timing comparisons, so the claimed efficiency and accuracy are not established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The node sub-network's fixed dense-layer input makes the reported 24–384 element convergence study undefined; no variable-size mechanism is specified.","rationale":"The reader's weakest assumption correctly identifies the variable-size input problem as the central technical gap. The paper's own text confirms that the node sub-network uses dense layers and that the experiments sweep mesh sizes from 24 to 384 elements, but it never explains how a fixed-dimension dense layer accommodates inputs of different lengths. This is not a stylistic omission; it directly affects whether the convergence study in Figure 3 is even well-defined. If the authors used separate networks per mesh, the 'surrogate model' claim becomes a family of models, and the comparison across resolutions is not a convergence study of a single method. If they used a shared network, a critical architectural detail is missing, making the work irreproducible. Other concerns, such as the undefined material penalty loss and the absence of runtime measurements, are also valid, but they are secondary: the variable-size architecture issue must be resolved before any of the numerical results can be trusted. The recommended verdict remains REJECT, consistent with the reader's assessment, because the concern is load-bearing and currently unaddressed.","tokens_in":14643,"tokens_out":4297,"duration_ms":45415,"concrete_test":"Obtain or reimplement the forward pass from Section 3, then instantiate the node sub-network and feed it two input vectors: one for a 24-element mesh (e.g., 25 nodal coordinates plus boundary flags) and one for a 384-element mesh (385 coordinates). If the first dense layer's weight matrix has a fixed column count, one of these forward passes will fail with a shape mismatch, proving that a single model cannot handle both meshes. Then check whether the authors' training code trains one model per mesh size or one shared model; if it trains one per mesh size, re-run the Figure 3 convergence study with a single model across all mesh sizes to see whether the reported error trends change.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of accurate displacement prediction across varying mesh resolutions (Section 5) rests on experiments in Section 4 that report H1 errors at 24, 48, 96, 192, and 384 elements. Section 3 describes the node sub-network as taking nodal coordinates, boundary conditions, and geometric factors into 'a series of dense layers.' Dense layers have fixed input dimensions, so a single network cannot accept inputs of different lengths unless an encoding mechanism (padding, masking, set aggregation, weight sharing, or separate models per mesh count) is present. No such mechanism is described anywhere in the manuscript. If separate models were trained per mesh count, then the '80-sample training' and the convergence curves in Figure 3 compare different models, and the idea of a single mesh-flexible surrogate is unsupported. If a single model was used, the input encoding is undefined and the results are not reproducible. Either way, the load-bearing experimental evidence for the headline accuracy and efficiency claims cannot be interpreted as stated. This gap is more fundamental than the missing baseline or runtime benchmarks because it undermines the validity of all reported numerical results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hybrid surrogate model that combines a one-dimensional Virtual Element Method (VEM) formulation for Euler–Bernoulli beams with a deep neural network split into node-specific and material-specific sub-networks. Training uses Sobolev derivative information and GradNorm loss balancing. The paper reports H1 errors for quadratic (order 4) and cubic (order 5) VEM formulations on a portico geometry across mesh sizes from 24 to 384 elements, using an 80-sample training set, and claims that the approach achieves accurate displacement predictions with very little data and faster inference than traditional numerical methods.","tokens_in":14819,"tokens_out":5250,"duration_ms":52355,"significance":"The idea of pairing a VEM solver with a neural-network surrogate for parametric beam analysis is potentially useful, particularly because the reported training set is small and the architecture separates nodal from material features. The manuscript also gives a clear, compact derivation of the VEM beam formulation and includes a pseudocode description of Sobolev training with GradNorm. However, as written, the central numerical claims cannot be checked: the primary result figure is absent, the material penalty loss L3 is never defined, the handling of variable mesh sizes is unspecified, validation is performed only against the same VEM generator that produced the training data, and the advertised inference-time advantage is never measured. At this stage the paper does not establish its main contributions.","major_comments":[{"comment":"The H1-error results that constitute the paper's central numerical evidence appear only as a reference to Figure 3, but no figure or data table with the mean and standard deviation values is included in the manuscript. Without the actual numbers or plots, the claimed convergence behavior and the differences between the order 4 and order 5 models cannot be verified or reproduced.","section":"§4, Figure 3"},{"comment":"The 'material penalty loss' L3, which appears in the total loss Ltotal = θ1L1 + θ2L2 + θ3L3 and in the GradNorm gradient norms, is never given a concrete definition. Algorithm 1 states only 'Compute the material penalty loss: L3 (problem-specific regularization)' and Section 3.3 says it 'penalizes deviations in material properties,' but no functional form, target quantity, or weighting is specified, making the training procedure irreproducible.","section":"§3.3, Eq. (39) and Algorithm 1"},{"comment":"The node sub-network is described as a series of dense layers taking nodal coordinates, boundary conditions, and geometric factors as inputs, yet the experiments sweep 24, 48, 96, 192, and 384 elements. Dense layers require fixed input dimensions, and the manuscript never states whether separate models are trained per mesh count or whether a variable-size encoding (padding, masking, aggregation, weight sharing) is used. Either way, the reported convergence study and the claim of a mesh-flexible surrogate are not interpretable as stated.","section":"§3 and §4"},{"comment":"The conclusion lists 'the efficiency of performing inference compared to traditional methods' among the demonstrated strengths, but Section 4 contains no timing measurements, no runtime comparison with VEM or FEM, and no complexity analysis. The efficiency claim is therefore unsupported by any experimental evidence in the manuscript.","section":"§5 and §4.2"},{"comment":"All reported H1 errors are computed against 'reference solutions obtained using the VEM formulation' that also generated the 80 training samples, and no analytical beam solution or independent solver is used for comparison. Because the training and test targets come from the same VEM implementation, the reported accuracy cannot distinguish genuine generalization error from systematic errors in the data generator. In addition, the unexpected increase in error at 384 elements is reported without a diagnostic analysis; Section 4.1 offers only qualitative speculation about nonlinearity, which does not establish the cause or demonstrate that the result is not an artifact.","section":"§4"}],"minor_comments":[{"comment":"The captions for Figures 1 and 2 appear without the corresponding images, and Figure 1's caption literally reads 'Sample figure caption,' indicating an unresolved placeholder.","section":"Figures 1 and 2"},{"comment":"The terms 'quadratic (order 4)' and 'cubic (order 5)' should be defined in relation to the polynomial degree n used in the VEM formulation of Section 2, because the current terminology is ambiguous.","section":"§4"},{"comment":"The abstract and introduction frame the contribution as solving one-dimensional Euler–Bernoulli beams, while Section 4 evaluates a portico (two-dimensional frame) geometry; this scope mismatch should be clarified.","section":"Abstract and §4"},{"comment":"Equation (42) writes the square of the L2 norm as equal to the L2 norm itself, and Eq. (41) then combines these quantities in a way that is dimensionally inconsistent; the standard definitions of \\(\\|u-u_h\\|_{L^2}^2\\) and \\(\\|\\nabla(u-u_h)\\|_{L^2}^2\\) should be used.","section":"Appendix A, Eq. (41)–(42)"},{"comment":"There is a typo in 'Physical-Informed Neural Networks' in the paragraph after the description of PINNs; it should read 'Physics-Informed.'","section":"§1"}],"recommendation":"reject","confidential_remarks":"The manuscript is at an early stage and lacks the essential experimental details expected of a journal submission: the primary figure is absent, the loss function L3 is undefined, the handling of variable-size meshes is unspecified, and validation is self-referential. If the authors add the missing definitions, figures, independent validation, and runtime measurements, a substantially revised submission could be worth considering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an early-stage hybrid VEM–deep learning surrogate for Euler-Bernoulli beams. The building blocks are all prior work—VEM for beams, Sobolev training, GradNorm—and the only new piece is splitting the network into node and material sub-networks. That split is a reasonable idea, and the paper is honest that the framework is preliminary. It also flags the unexpected H1 error increase at fine meshes and discusses possible causes; that is appropriate, not concealment.\n\nWhat the paper does well: the VEM formulation section is competent and the architecture description is clear. Using Sobolev training to enforce smoothness and GradNorm to balance three losses is a sensible combination for this problem. Training on 80 samples and reporting errors across 24–384 elements shows they did run experiments, and the claim that few samples suffice is at least plausible.\n\nWhere it falls apart: the experimental evidence cannot be checked. The material penalty loss L3 is never defined—Algorithm 1 just calls it 'problem-specific regularization.' The H1 errors are described but the actual figures are not in the text, so we can't see the convergence curves. There is no baseline against an analytical solution or an independent FEM implementation; the target is always the same VEM code, so the reported errors only measure fit to the internal generator. The claim of faster inference is never measured. And the stress-test concern is real: the node sub-network is described as dense layers, which have fixed input size, yet the paper reports convergence studies over meshes with 24 to 384 elements. Unless they trained separate models per mesh count or used some encoding, the results are undefined. The paper gives no mechanism. That undermines every reported number.\n\nOn the flip side, some of the reader's concerns are slightly overstated. The paper never claims to be a new mathematical result; it's explicitly a surrogate-modeling exploration. The absence of code is a reproducibility problem, not a sign of deception. And the literature review is fine—citations are appropriate.\n\nBottom line: this is a workshop-level idea with a plausible architecture and an honest discussion of limitations, but the load-bearing evidence is missing. The right next step is a revised version with open code, a defined L3, a variable-mesh encoding or separate per-mesh models, and at least one external baseline. As is, I wouldn't send it to peer review; it needs to be reworked into something checkable first.","headline":"A plausible but underspecified neural-VEM surrogate whose headline claims can't be checked because key losses and the variable-mesh mechanism are missing.","tokens_in":15412,"tokens_out":4577,"would_cite":false,"duration_ms":36273,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","74K10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A VEM-trained two-branch neural network can predict beam displacement fields from just 80 samples, with errors that shrink under mesh refinement until an unexpected rise at 384 elements.","keywords":["Virtual Element Method","deep learning surrogate","Euler-Bernoulli beam","Sobolev training","GradNorm","displacement field prediction","structural mechanics","portico frame"],"falsifier":"Evaluate the trained network on a resolution not seen in training, such as 768 elements, and check whether it even accepts the input and whether the H1 error continues a smooth trend; if it cannot predict or the error jumps far off the VEM convergence curve, the claimed mesh-agnostic surrogate fails.","tokens_in":14371,"feed_emoji":"🏗️","tokens_out":9280,"duration_ms":78840,"temperature":0.7,"pith_summary":"This paper tries to establish that a surrogate model combining the Virtual Element Method (VEM) with a two-branch neural network can predict displacement fields of one-dimensional Euler-Bernoulli beams and portico frames across varying material and geometric parameters. The network keeps nodal geometry and material properties in separate sub-networks, then trains on only 80 VEM solutions using Sobolev training, which matches derivatives, and GradNorm, which balances the loss terms. Reported H1 errors against VEM references generally decrease as the mesh refines, with the order-4 model slightly outperforming the order-5 model, and both models showing an unexpected error increase at 384 elements. If the claim holds, structural surrogates could be built from small datasets and evaluated quickly enough for iterative design and analysis.","feed_headline":"Surrogate trained on 80 samples predicts beam displacements","feed_subtitle":"Training needs only 80 VEM solutions and inference is fast enough for iterative design and analysis.","key_machinery":"The carrying mechanism is the two-branch neural network: a node sub-network of dense layers receives coordinates, boundary conditions, and geometry while a parallel material sub-network receives Young's modulus, area, and moment of inertia, and the concatenated outputs pass through final dense layers to predict the deflection field. Training is driven by a composite loss of displacement error, a Sobolev term penalizing mismatches in projected derivatives, and a material-penalty term, with the three contributions balanced adaptively by GradNorm. The VEM side supplies the reference solutions through a general-order projection formulation that introduces internal moment degrees of freedom for higher-order beam elements.","core_discovery":"On its own terms, the paper's central discovery is that a neural network can learn the parameter-to-displacement map of Euler-Bernoulli beam problems from VEM-generated examples, provided the architecture separates node data from material data and the training loss includes derivative information with dynamically balanced task weights. The experiments show a mean H1 error that decreases from 24 to 192 elements, indicating the learned surrogate tracks VEM's refinement behavior, and that 80 training samples suffice for this tracking. The paper also discovers that the tracking fails at 384 elements, where the error rises for both quadratic and cubic formulations, a deviation it attributes to interference between the network's learned non-linear approximation space and VEM's expected polynomial convergence.","pith_inferences":["The paper leaves open how a fixed-size dense node sub-network accommodates element counts from 24 to 384; an editor-level inference is that a genuinely mesh-invariant encoding, such as aggregated per-element features, is needed before the convergence results can be read as a single model.","The plateau and rise at fine meshes suggest network capacity, not VEM order, is the bottleneck; scaling network width or depth with element order is a direct testable extension.","Because material parameters enter as global scalars, the same architecture could plausibly predict at extrapolated E, A, and I values, but the paper does not test extrapolation, so that experiment remains open.","The hybrid recipe could extend to two- and three-dimensional elasticity if the node branch is replaced with a geometrically aware encoder, since the material branch would still operate on scalars."],"forward_implications":["If the surrogate generalizes as reported, structural analysis can precompute a small set of VEM solutions and then run fast inference for other parameter values.","The two-branch design is a reusable pattern for problems where one part of the input is geometric or nodal and another part is scalar material data.","The Sobolev-plus-GradNorm training recipe is presented as a way to keep multi-term physics losses stable when training data is scarce.","Fast inference would support real-time design iteration, optimization, and uncertainty quantification in one-dimensional frame analysis.","The error rise at 384 elements implies the current hybrid does not inherit VEM's asymptotic convergence and needs architectural changes before it can be trusted on fine meshes."],"supporting_citations":[{"why":"Foundational VEM paper that supplies the projection-based discretization philosophy this method builds on.","marker":"[1]"},{"why":"Virtual element formulation for trusses and beams that the one-dimensional beam and portico formulation extends.","marker":"[9]"},{"why":"Sobolev training technique that provides the derivative-matching loss used to enforce smoothness.","marker":"[27]"},{"why":"GradNorm algorithm that supplies the dynamic loss-weighting mechanism for balanced multi-task training.","marker":"[28]"},{"why":"Deep learned finite elements, a prior hybrid DL-FEM approach that the proposed surrogate is positioned against.","marker":"[22]"},{"why":"FEM-enhanced neural networks for forward and inverse problems, another hybrid baseline whose architecture is contrasted with the two-branch split.","marker":"[26]"}],"fun_headline_variants":["80 VEM samples train a net for beam displacement","Hybrid VEM-DL predicts beams with 80 examples","Neural beam surrogate tracks VEM, then fails at 384","VEM-deep hybrid learns beam displacement from few samples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume one dense node sub-network can handle meshes from 24 to 384 elements, but no mechanism is stated for how a fixed-size network input represents a variable-size mesh.","fun_headline_variants_meta":{"raw":{"variants":["80 VEM samples train a net for beam displacement","Hybrid VEM-DL predicts beams with 80 examples","Neural beam surrogate tracks VEM, then fails at 384","VEM-deep hybrid learns beam displacement from few samples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00119,"raw_usage":{"total_tokens":4867,"prompt_tokens":856,"completion_tokens":4011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":3942}},"tokens_in":472,"tokens_out":4011,"duration_ms":28455,"temperature":1.0,"reasoning_tokens":3942,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:58.651730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained network on a resolution not seen in training, such as 768 elements, and check whether it even accepts the input and whether the H1 error continues a smooth trend; if it cannot predict or the error jumps far off the VEM convergence curve, the claimed mesh-agnostic surrogate fails.","supporting_citations":[{"cited_title":"Beirão da Veiga, F","cited_arxiv_id":null,"evidence_quote":"Foundational VEM paper that supplies the projection-based discretization philosophy this method builds on."},{"cited_title":"Wriggers","cited_arxiv_id":null,"evidence_quote":"Virtual element formulation for trusses and beams that the one-dimensional beam and portico formulation extends."}],"review_version":1}