{"id":"6f5aebde-d883-49a6-b40d-208ec93da745","arxiv_id":"2411.09728","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural network predicts finite element model errors and upscales coarse elastic plate solutions, but the claimed benefit of the physics losses is undermined because the displacement loss duplicates the main error loss.","lead":"This paper trains a neural network to predict the difference between coarse and fine finite element simulations of a stretched plate with a hole, and to upscale coarse outputs to fine mesh resolution. The authors claim physics-informed loss terms beat pure data-driven training, but the main physics term is algebraically the same as the baseline error loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7) Lu is algebraically identical to beta_1 times Eq. (6) Lerror; the reported physics-loss improvement is a loss-reweighting artifact, not evidence of a physics-informed gain.","rationale":"The most load-bearing concern is the algebraic identity between Lu and Lerror, because it invalidates the only experimental evidence for the central claim that physics-informed losses outperform a purely data-driven approach. The reader's stated weakest_assumption, the convergence of the Q8 ground truth, is an external-validity concern that would affect whether the learned error is the 'true' model error, but it does not affect whether the comparative experiment can distinguish physics-informed from data-driven learning at all. The algebraic identity is an internal inconsistency, making it more fundamental. The reader's rationale does mention this identity, but the formal weakest_assumption field points to Q8 convergence, so I disagree that this is the single most load-bearing concern. I concur with the reader's REJECT verdict, and I would keep it unchanged: the basic error-prediction demonstration may be viable, but the paper's comparative claim about physics-informed losses is unsupported by its own equations.","tokens_in":11447,"tokens_out":4811,"duration_ms":44673,"concrete_test":"At any training checkpoint, compute the parameter gradients of Lu and Lerror. If grad_theta Lu = beta_1 * grad_theta Lerror, then Lu imposes no independent constraint. Alternatively, retrain Case 2 with beta_1 fixed to 1 and Case 3 with the same seed and data order; if per-sample losses and test Lerror coincide up to floating-point precision, the physics-informed loss term is redundant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim is that physics-informed loss functions outperform a purely data-driven approach. Eq. (7) defines Lu = (beta_1/N) sum || (e^A_n + u_R,n) - u_(H,Omega_Q4),n ||_1. By Eq. (1), the true model error is e^d_n = u_(H,Omega_Q4),n - u_R,n. Substituting, each term of Lu becomes || e^A_n - e^d_n ||_1, which is exactly the term in Lerror (Eq. (6)). Hence Lu = beta_1 * Lerror. The total loss in Case 2 is (1 + beta_1) Lerror, a scaled version of the purely data-driven Case 3 loss. A learnable beta_1 cannot add information; it only rescales the gradient. Therefore the reported reduction in testing loss (Case 2: 1.14e-6 vs Case 3: 1.23e-6, Table 1) cannot be attributed to physics, especially since the difference is smaller than the reported standard deviations (2.90e-6 and 2.30e-6). This internal inconsistency directly undermines the paper's headline claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a neural-network-based explicit model error approximator for finite element solutions. The network takes reduced-order (Q4) displacement fields as input and outputs both a predicted model error field (the Q4-to-Q8 difference) and a Q8 displacement field for superresolution. The training loss combines an error term, a displacement loss, and a superresolution loss, with the last two described as physics-informed. Using a synthetic dataset of 10,000 elastic plate simulations, the authors compare three loss configurations and report that including the displacement loss improves testing accuracy over purely data-driven training. The paper also presents dropout-based uncertainty estimates and qualitative superresolution results.","tokens_in":11729,"tokens_out":5471,"duration_ms":55711,"significance":"If correct, explicit machine-learned error approximators that correct coarse finite element solutions toward finer ones would be practically useful, and the paper's dataset-generation pipeline is a reasonable test bed. The basic error-prediction results look plausible: prediction-difference histograms are centered near zero and mean absolute differences are below 4e-6 m. However, the central comparative claim is not supported: the displacement loss is algebraically identical to the error loss, so the reported 'physics-informed' gain is a reweighting artifact. The paper also does not verify that the Q8 solution is a converged ground truth and does not provide code. These issues undermine the stated novelty and the title/abstract claims.","major_comments":[{"comment":"Equation (7) defines Lu = (β1/N) Σ ||(e^A_n + u_R,n) − u_(H,Ω_Q4),n||_1. Since the true model error in Eq. (1) is e^d_n = u_(H,Ω_Q4),n − u_R,n, each term of Lu equals ||e^A_n − e^d_n||_1, making Lu = β1·Lerror. Consequently the total loss in Case 2 is (1 + β1)·Lerror, a scaled version of the Case 3 loss. A learnable β1 cannot inject independent physical information; it only rescales the gradient magnitude. The reported improvement between Cases 2 and 3 (testing loss 1.14 vs 1.23 × 10^-6 m) is smaller than the reported standard deviations (2.90 and 2.30 × 10^-6 m) and therefore cannot be attributed to physics. This invalidates the headline claim that physics-informed loss functions outperform a purely data-driven approach.","section":"Loss Functions and Model Training"},{"comment":"The Q8 discretization is designated as the 'higher-order model' and its solution is used as ground truth in Eq. (1), but no convergence study is provided. Q4 and Q8 differ simultaneously in polynomial order and mesh density (861 vs 9,841 nodes). Without demonstrating that the Q8 solution is effectively converged with respect to h- and p-refinement, the learned target is only a difference between two imperfect discretizations, and correcting Q4 toward Q8 does not necessarily reduce the actual numerical error. This premise is load-bearing because the entire supervised objective is defined by this difference.","section":"Numerical Simulation and Dataset Generation"},{"comment":"The term 'physics-informed' is used for losses that are not physics-based. Lsuper in Eq. (8) is a standard supervised loss comparing a network output to Q8 nodal values; it imposes no PDE, equilibrium, or constitutive constraint. Lu, as shown above, reduces to the data-fidelity error loss. No physical residual is used anywhere in the network. Thus the comparison in Table 1 does not isolate any physics-informed mechanism, and the description of the method as a PINN is misleading.","section":"Loss Functions and Model Training"},{"comment":"The claim that the method achieves superresolution in both the h and p senses is not supported. The network maps Q4 nodal displacements to Q8 nodal displacements, which is an h-type upsampling (more nodes). No mechanism changes the element interpolating polynomial basis, so the claimed p-refinement novelty is not demonstrated. The superresolution results are also only shown qualitatively for one sample, without a quantitative error metric on the testing set.","section":"Superresolution"}],"minor_comments":[{"comment":"There are duplicated phrases in the text: 'conducted using using' and 'outputs from the from the numerical simulation' should be corrected.","section":"Numerical Simulation"},{"comment":"Table 1 has formatting issues ('1 .38') and does not specify whether the reported standard deviations are over epochs, over seeds, or over testing samples; the testing standard deviations are larger than the mean differences used to support the physics-informed gain, so the uncertainty should be quantified more carefully.","section":"Table 1"},{"comment":"The notation for the corrected solution, written as 'e_u_H', is undefined and appears to be a typo; it should be clearly distinguished from the true higher-order solution u_H.","section":"Formulation"},{"comment":"The data availability statement says data are available upon reasonable request, but no code or trained model is provided; releasing code would substantially improve reproducibility, especially given the difficulty of reproducing the exact loss-weighting scheme.","section":"Data Availability Statement"},{"comment":"The learned values of the loss coefficients β1 and β2 are never reported, so the reader cannot evaluate the effective weighting of the loss terms or compare the results across configurations.","section":"Loss Functions"}],"recommendation":"reject","confidential_remarks":"The central claim of the paper — that physics-informed loss functions enable NNs to surpass data-driven approaches — is invalid because the displacement loss is algebraically identical to the error loss. The paper could potentially be reframed as a data-driven error-correction and superresolution study, but that would require changing the title, abstract, and conclusions, and the original claim would need a genuinely physics-based loss to be supported. This is a fundamental issue rather than a local correction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a useful-and-true core (a two-branch NN that predicts Q4-to-Q8 model error and upsamples displacement), but its headline claim is not supported by its own equations. Eq. (7) defines Lu as beta_1/N times the sum of ||(e^A_n + u_R,n) - u_(H,Omega_Q4),n||_1. Since the true error e^d_n = u_(H,Omega_Q4),n - u_R,n, each term inside Lu is exactly the corresponding term in Lerror (Eq. (6)). So Lu = beta_1 * Lerror. In Case 2 the total loss is (1 + beta_1) Lerror, a scaled version of Case 3. A learnable beta_1 cannot add information; it only rescales the gradient. The reported improvement in testing Lerror (1.14 vs 1.23e-6 m) is also smaller than the reported standard deviations (2.90 and 2.30e-6), so it is not even a meaningful numerical difference. The claim in the abstract and conclusions that physics-informed loss functions enable NNs to surpass a purely data-driven approach therefore fails.\n\nWhat is actually new: the multi-task architecture with a Q8 superresolution head, and the framing of p-refinement in addition to h-refinement, is a legitimate extension of earlier work (Smyl et al. 2021; Freno and Carlberg 2019). The basic error-prediction demonstration is plausible: the prediction histograms are centered near zero and the mean absolute difference is below 4e-6 m. The MC-dropout uncertainty analysis is a nice addition.\n\nWhere it is soft: besides the load-bearing loss identity, there is no baseline comparison to prior error-correction or superresolution methods, so we don't know whether the architecture adds value over simpler baselines. The Q8 solution is treated as ground truth without checking mesh convergence; that is a standard assumption for a demonstration, but it limits the claim that they are approximating 'the' model error. Data and code are not released, only available on request, which is a reproducibility weakness for an empirical paper.\n\nWho this is for: researchers in computational mechanics and inverse problems who want explicit nodal error maps and a cheap way to upgrade coarse FE solutions. The basic idea is worth discussing, but the paper as written overclaims. A serious referee should see it, because the error is fixable by reframing the contribution away from 'physics-informed beats data-driven.' I would not cite it in its current form.\n\nRecommendation: send it to peer review, but tell the authors the physics-informed comparison needs to be dropped or replaced with a real independent physics constraint; otherwise the paper is just a scaled data-driven model.","headline":"The paper's core ablation is invalid: Eq. (7) is algebraically identical to beta_1 times Eq. (6), so the claimed physics-informed improvement is just loss reweighting.","tokens_in":12268,"tokens_out":2673,"would_cite":false,"duration_ms":32691,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Physics-informed loss terms let a neural network approximate the nodal difference between coarse and fine finite-element solutions and output the fine mesh's displacement field, beating a purely data-driven baseline.","keywords":["physics-informed neural networks","model error approximation","finite element analysis","superresolution","Q4/Q8 discretization","displacement fields","uncertainty quantification","elastic plate with opening"],"falsifier":"Take several test samples and compare the Q8 solution against a substantially finer reference mesh (for example, 40,000 or more Q8 nodes) or against an analytical solution. If the per-node difference between Q8 and that reference is comparable to or larger than the PINN's predicted model error, then correcting Q4 toward Q8 cannot reduce the true error, and the claim that the network approximates numerical model error would be falsified.","tokens_in":11246,"feed_emoji":"📐","tokens_out":9701,"duration_ms":83265,"temperature":0.7,"pith_summary":"The paper sets out to show that a neural network can explicitly approximate numerical model errors in finite element analysis—the gap between a coarse four-node quadrilateral (Q4) mesh solution and a fine eight-node quadrilateral (Q8) mesh solution—rather than only supplying a scalar error bound or a corrected operator. The proposed physics-informed neural network (PINN) takes the Q4 displacement field as input and outputs both the model-error field at Q4 nodes and the Q8 displacement field, effectively combining error approximation with superresolution. Trained on 9,000 randomized samples of a two-dimensional elastic plate with a central opening and tested on 1,000 held-out samples, the network reproduces the error fields closely, with mean absolute differences below $4.0\\times10^{-6}$ m in most areas. The central comparison is an ablation showing that adding a displacement-consistency physics loss reduces mean testing error by 7.3% relative to the purely data-driven setup, which is the evidence for the paper's claim that physics-informed loss functions let networks surpass data-only approximation of model errors.","feed_headline":"Physics loss helps neural nets correct finite-element errors","feed_subtitle":"Two physics-based loss terms lift Q4 meshes toward Q8 accuracy and beat data-only training.","key_machinery":"The central object is a two-branch physics-informed neural network whose input is the Q4 nodal displacement vector (1,722 entries). One branch outputs predicted model error at Q4 nodes (1,722 entries); the other outputs a predicted Q8 displacement field (19,682 entries), which is the superresolution branch. Training minimizes a weighted total loss $\\mathcal{L}_{\\text{total}} = \\mathcal{L}_{\\text{error}} + \\beta_1 \\mathcal{L}_u + \\beta_2 \\mathcal{L}_{\\text{super}}$, where $\\mathcal{L}_u$ is the displacement-consistency term that penalizes deviations of the corrected Q4 field from the Q8 solution sampled at Q4 nodes, and $\\mathcal{L}_{\\text{super}}$ directly supervises the Q8 output. The $\\mathcal{L}_u$ term is the load-bearing physics-informed component: the ablation shows it reduces both training (12.0%) and testing (7.3%) model-error loss, while $\\mathcal{L}_{\\text{super}}$ adds the higher-resolution output. Dropout is kept active at test time to approximate predictive uncertainty.","core_discovery":"The central claim is that the learned error approximator $A(\\Theta)$ in $\\tilde{u}_H \\approx u_R + A(\\Theta)$ captures the numerical model error $e = u_H - u_R$ closely enough that correcting a Q4 solution moves it toward Q8 accuracy, and that the same network can directly emit the Q8 displacement field. The support comes from a two-branch network trained with three loss terms: $\\mathcal{L}_{\\text{error}}$ for the nodal error, $\\mathcal{L}_u$ enforcing that error-plus-Q4-input match the Q8 displacement sampled on the Q4 mesh, and $\\mathcal{L}_{\\text{super}}$ enforcing that the second branch match the full Q8 displacement. Removing $\\mathcal{L}_u$ raises mean testing loss by 7.3%, and removing both physics terms raises it further, so the physics constraints contribute beyond the data-driven mapping. Dropout-based uncertainty places prediction spread at $2{-}2.5\\times10^{-6}$ m, well below the model-error magnitudes, while the superresolution branch produces Q8 fields directly from Q4 inputs.","pith_inferences":["A direct testable extension is to hold out a third, even finer reference solution (or an analytical solution) and ask whether the corrected Q4 field moves toward that reference; if it does not, the learned quantity is mesh-to-mesh difference rather than true model error.","The correction formula $\\tilde{u}_H \\approx u_R + A(\\Theta)$ could be applied recursively: feed the corrected field back as input to approximate the error at the next refinement level, enabling multi-level error correction without re-running the fine solver.","Because the input is the raw Q4 nodal field, the same two-branch architecture could be retrained for other mesh pairs (for example, linear-to-quadratic or triangular-to-quadrilateral elements) if corresponding randomized datasets are generated."],"forward_implications":["Coarse Q4 solutions can be corrected to Q8-level nodal accuracy with a single forward pass, giving engineers explicit spatial error maps instead of scalar error bounds.","Adding the displacement-consistency physics loss ($\\mathcal{L}_u$) is shown to cut testing model-error loss by 7.3% over the data-only baseline, so physics-informed regularization provides a measurable accuracy gain.","The dual output makes the network a superresolver in both mesh fineness ($h$) and interpolation order ($p$), producing Q8 displacement fields from Q4 inputs.","Dropout uncertainty estimates in the $2{-}2.5\\times10^{-6}$ m range give a practical confidence measure for the predicted error field."],"supporting_citations":[{"why":"Introduces the PINN framework whose physics-loss training the paper adapts from PDE solving to model-error approximation.","marker":"Raissi et al. 2019"},{"why":"Establishes the task of learning and correcting non-Gaussian model errors and supplies the choice of reduced-order solution as network input.","marker":"Smyl et al. 2021"},{"why":"Formulates machine-learning error models for approximate parameterized solutions, the baseline approach this work extends to nodal error fields.","marker":"Freno and Carlberg 2019"},{"why":"Provides the dropout-as-Bayesian-approximation method used to quantify prediction uncertainty.","marker":"Gal and Ghahramani 2016"},{"why":"Supplies the finite element code adapted to generate the randomized Q4/Q8 training dataset.","marker":"Krysl 2010"},{"why":"Presents the Bayesian approximation error framework that is contrasted with explicit learned error approximators.","marker":"Kaipio and Somersalo 2006"}],"fun_headline_variants":["PINNs correct FEM errors and boost mesh resolution","Physics-informed nets fix finite-element model errors","Error-correcting PINNs: from Q4 mesh to Q8 accuracy","Physics loss guides neural nets to correct FEM errors","Neural nets learn FEM error + superresolution in one go"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Q8 solution at 9,841 nodes is the true higher-order answer; if that mesh is itself not converged, the network is learning the difference between two imperfect meshes rather than the actual numerical model error.","fun_headline_variants_meta":{"raw":{"variants":["PINNs correct FEM errors and boost mesh resolution","Physics-informed nets fix finite-element model errors","Error-correcting PINNs: from Q4 mesh to Q8 accuracy","Physics loss guides neural nets to correct FEM errors","Neural nets learn FEM error + superresolution in one go"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2773,"prompt_tokens":954,"completion_tokens":1819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1740}},"tokens_in":570,"tokens_out":1819,"duration_ms":12954,"temperature":1.0,"reasoning_tokens":1740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:28:33.598906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take several test samples and compare the Q8 solution against a substantially finer reference mesh (for example, 40,000 or more Q8 nodes) or against an analytical solution. If the per-node difference between Q8 and that reference is comparable to or larger than the PINN's predicted model error, then correcting Q4 toward Q8 cannot reduce the true error, and the claim that the network approximates numerical model error would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the PINN framework whose physics-loss training the paper adapts from PDE solving to model-error approximation."},{"cited_title":"N., Black, J","cited_arxiv_id":null,"evidence_quote":"Establishes the task of learning and correcting non-Gaussian model errors and supplies the choice of reduced-order solution as network input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formulates machine-learning error models for approximate parameterized solutions, the baseline approach this work extends to nodal error fields."},{"cited_title":"and Ghahramani, Z","cited_arxiv_id":null,"evidence_quote":"Provides the dropout-as-Bayesian-approximation method used to quantify prediction uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the finite element code adapted to generate the randomized Q4/Q8 training dataset."},{"cited_title":"and Somersalo, E","cited_arxiv_id":null,"evidence_quote":"Presents the Bayesian approximation error framework that is contrasted with explicit learned error approximators."}],"review_version":1}