{"id":"63ea343c-5ae3-4fd0-85f7-56e775a49f4f","arxiv_id":"2508.00029","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A quantum-classical MLP is claimed to reach MSE 3.16e-11 on bridge inverse FE displacement prediction, but the test-set statistics in the paper are mutually inconsistent and no code or data is provided.","lead":"The paper proposes a hybrid quantum-classical network to predict full bridge displacement fields from seven sensor inputs, reporting a very low mean squared error of 3.16e-11. The result matters because fast inverse finite element surrogates could update digital twins in near real time, but the reported metrics are internally inconsistent and the code is absent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 metrics are algebraically inconsistent on a single test set: implied target variances differ by ~7 orders of magnitude, so the claimed large-margin QMLP advantage is unsupported.","rationale":"The paper's key empirical assertion is that QMLP reaches MSE 3.16e-11 and outperforms classical baselines by a large margin. For that assertion to be supported, Table 1 must be a set of metrics computed on one test set. The arithmetic of the table falsifies this premise: using the standard identities R2 = 1 - MSE/Var(y_test) and NRMSE(std) = RMSE/std(y_test), different rows imply target variances that differ by seven orders of magnitude, and the R2 ranking contradicts the MSE ranking. I verified this with the printed numbers; it is an internal consistency failure, not a disagreement with any external consensus. The only way the central claim could survive is if the metrics were computed under different, unstated target scalings or test splits, in which case the comparison is not the apples-to-apples comparison the text claims. The reader's REJECT verdict is therefore appropriate, and no additional concern is needed. The manuscript provides no code, data, or formal verification that could independently corroborate the table, so keeping the reader's verdict unchanged is the honest assessment.","tokens_in":12214,"tokens_out":7066,"duration_ms":79001,"concrete_test":"Obtain the original saved predictions and the identical test target vector y_test for all six models, and run one script that computes MSE, RMSE, R2 = 1 - MSE/Var(y_test), NRMSE(range) = RMSE/(max(y_test) - min(y_test)), and NRMSE(std) = RMSE/std(y_test). Verify that Var(y_test) and std(y_test) are the same in every row and that the printed values match. If predictions/y_test are unavailable, perform the arithmetic check on Table 1 alone: deriving Var from each row yields at least three mutually incompatible values (about 5.95e-2, 2.19e-9, and 6.07e-8), so the published table cannot support the claimed advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All rows of Table 1 are claimed to be measured on the same final test set, so target variance V and standard deviation s must be common across rows. They are not. From the baseline row, MSE = 1.008e-3 and R2 = 0.9831 imply V = MSE/(1 - R2) = 5.95e-2 (s = 0.244). From the best QMLP row, MSE = 3.16e-11 and R2 = 0.9856 imply V = 2.19e-9 (s = 4.68e-5) — seven orders of magnitude smaller. The NRMSE columns give yet another set: baseline RMSE = 3.1e-2 with NRMSE(std) = 0.03513 implies s = 0.882, while QMLP RMSE = 5.62e-6 with NRMSE(std) = 0.00726 implies s = 7.74e-4, and Classic MLP (Clustering) with RMSE = 2.72e-5 and NRMSE(std) = 0.03513 also implies s = 7.74e-4. No single test set simultaneously has all these variances. The R2 ordering is also impossible on identical targets: QMLP (Poly-SPD + HC + clustering) has roughly 20x lower MSE than Classic MLP (Clustering) but lower R2 (0.9856 vs 0.9878). The abstract reports MSE 3.16e-11 while the contributions bullet states 3.1e-4. Without code or data, the central claim of large-margin QMLP superiority rests on internally inconsistent numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid quantum-classical multilayer perceptron (QMLP) for real-time inverse finite element modeling in a digital-twin setting. Seven tilt-meter readings are mapped to 1017 nodal displacements via a pipeline that combines polynomial feature expansion, symmetric positive definite (SPD) matrix construction, Hilbert-Schmidt vectorization, angle embedding into a parameterized quantum circuit, and a classical readout network. The authors report experiments on a bridge case study and claim that the best QMLP variant achieves an MSE of 3.16e-11, outperforming purely classical baselines by a large margin. The paper also includes a complexity comparison and a discussion of implications for structural health monitoring.","tokens_in":12622,"tokens_out":4030,"duration_ms":47419,"significance":"If the empirical results were reliable and reproducible, the paper would offer a useful application-oriented demonstration of a quantum-classical surrogate for inverse FE modeling, with a clearly described architectural pipeline and a rough complexity estimate. The methodology combines several known techniques (polynomial features, SPD encoding, variational circuits) in a plausible way for a low-dimensional-input, high-dimensional-output regression task. However, the central empirical claim rests entirely on Table 1, and that table is internally inconsistent. The reported MSE, RMSE, R2, and NRMSE values cannot all hold on the same test set, and the abstract's headline MSE contradicts the contributions section. Because the main evidence is not self-consistent, the paper does not currently support its central claim of large-margin QMLP superiority. The paper also provides no code or data, and the evaluation protocol is described too loosely to allow independent verification.","major_comments":[{"comment":"The metrics in Table 1 are mutually inconsistent under the stated caption that all metrics were measured on the final test set. For the QMLP (Poly-SPD + HC + clustering) row, MSE=3.16e-11 and R2=0.9856 imply, via R2 = 1 - MSE/Var(target), a target variance of 2.19e-9 (standard deviation 4.68e-5), whereas NRMSE(std)=0.00726 with RMSE=5.62e-6 implies a target standard deviation of 7.74e-4. For the baseline classic MLP row, MSE=1.008e-3 and R2=0.9831 imply a target standard deviation of 0.244, while NRMSE(std)=0.03513 with RMSE=3.1e-2 implies a standard deviation of 0.883. No single test set can simultaneously satisfy these relations, so at least some of the reported metrics are miscalculated or were computed under different conditions. Since the abstract's headline result and the comparison against classical baselines rely on these numbers, the central empirical claim is unsupported.","section":"Table 1"},{"comment":"The abstract reports an MSE of 3.16e-11 for the proposed QMLP, while the third contributions bullet in the Introduction states 'superior performance (MSE 3.1 × 10−4)'. These two values differ by seven orders of magnitude. This is not a minor typo: it changes the claimed improvement over the baseline from about seven orders of magnitude to about one order of magnitude. The manuscript needs to state which number is correct and explain the discrepancy before the results can be interpreted.","section":"Table 1 and Abstract"},{"comment":"The evaluation protocol is underspecified and appears to involve model selection on the test set. The text says that the number of clusters k was chosen by comparing NRMSE and R2 for different k (Figure 4), and then 'all metrics were measured on the final test set' (Table 1). If the same data were used both to select k=7 and to report final performance, the reported metrics risk being optimistically biased. The paper should describe a proper train/validation/test split, state how many random seeds were used, and report variance or confidence intervals. Without this information, the performance differences in Table 1 cannot be assessed.","section":"Section 2.4 and Results, 'Clustering Analysis'"},{"comment":"The R2 ordering is impossible on the same target variable. For Classic MLP (Clustering enforced), MSE=7.40e-10 and R2=0.9878; for QMLP (Poly-SPD + HC + clustering), MSE=3.16e-11 and R2=0.9856. On a fixed test set with fixed target variance, a lower MSE must yield a higher or equal R2. The reported values violate this identity, confirming that the MSE and R2 columns are not computed consistently. This further undermines the comparison and points to a systematic issue in the evaluation.","section":"Table 1"}],"minor_comments":[{"comment":"The phrase 'thw proposed model' in the description of QMLP (Poly-SPD) + clustering contains a typo and should read 'the proposed model'.","section":"Table 1 caption"},{"comment":"The feature expansion notation is inconsistent: the input vector is denoted x in Section 1.2 and Section 2.1, but the polynomial expansion in Eq. (2) uses x while later text reuses x for the expanded vector. The notation should be cleaned up to avoid confusion between raw measurements and expanded features.","section":"Section 2.1, Eq. (6)-(8)"},{"comment":"The text refers to 'Figure 2' and 'Figure 3' for the circuit and clustering plots, but the figure captions are interleaved confusingly; the circuit figure is mentioned before its caption appears, and Figure 3 is referenced for clustering results inside a section titled 'Clustering Analysis'. The figure numbering and in-text references should be checked.","section":"Section 2.3"},{"comment":"Some references appear incomplete or have placeholder formatting, and the reference list contains several self-citations. The authors should verify that all bibliographic entries are complete and that the citation style is consistent.","section":"References"}],"recommendation":"reject","confidential_remarks":"The internal inconsistency of Table 1 is not a stylistic issue but a load-bearing failure: the paper's only empirical evidence for the claimed quantum advantage is arithmetically impossible on a single test set. Even a generous reading of the architecture does not rescue the results, because the reported numbers cannot be used to compare models. The abstract/contributions MSE discrepancy adds to the impression that the quantitative results have not been carefully checked. I would not recommend major revision unless the authors can supply corrected, consistent metrics, a clear evaluation protocol, and ideally code/data for reproducibility; as submitted, the paper does not meet the evidentiary bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real application idea, but its central empirical claim cannot be checked because Table 1 is arithmetically inconsistent, and no code or data are provided.\n\nWhat is new: the specific stack — polynomial features, SPD embedding, Hilbert–Schmidt vectorization, angle embedding, a PQC, and cluster-sized hidden layers — applied to inverse FE displacement prediction from sparse sensor readings on a bridge. I haven't seen that exact combination in the SHM literature, and the problem (mapping 7 sensor values to 1017 nodal displacements) is a legitimate one. The authors also include a complexity estimate and a limitations section that does not pretend scaling is solved.\n\nWhat the paper does well: the motivation is clearly written, the FE dataset generation is described in enough detail to be reproducible in principle, and the clustering-based sizing of the last hidden layer is a reasonable engineering heuristic. The complexity estimate (roughly 2.5x classical inference) is useful context for real-time deployment.\n\nNow the soft spots, and they are not minor. Table 1 reports all models evaluated on the same final test set. On a fixed target vector, MSE, RMSE, R², and NRMSE(std) are algebraically tied to one another. They are not even close to consistent here. The baseline row: MSE 1.008e-3 with R² 0.9831 implies target variance ~5.95e-2. The best QMLP row: MSE 3.16e-11 with R² 0.9856 implies target variance ~2.19e-9 — seven orders of magnitude smaller. The NRMSE(std) column gives yet another set of implied variances (0.882 for baseline, 7.74e-4 for the two clustering rows). No single target vector can satisfy all of these. Also, the Abstract states MSE 3.16e-11 while the contributions bullet in the body says MSE 3.1e-4. That is a two-order-of-magnitude discrepancy inside the same paper. Without code or data, the reader has no way to tell which number is real, and the central claim of large-margin QMLP superiority is unsupported.\n\nOne more thing: even taken at face value, the best QMLP's R² (0.9856) is lower than the classical clustered MLP's (0.9878). So the \"large margin\" depends on MSE alone, which is exactly the metric that is inconsistent.\n\nBottom line: the architecture is worth a look for someone working on quantum-classical surrogates for inverse problems, but the reported evidence does not support the paper's main conclusion. The right path is to fix the metrics, release code and data, and rerun the comparison on a consistent scale. As it stands, I would not send this to referees; it needs a major revision first.\n\nRecommendation: desk reject with an invitation to resubmit after a rigorous correction of the evaluation.","headline":"Interesting architecture for quantum-classical inverse FE surrogates, but the central empirical claim is unsupported by internally inconsistent metrics.","tokens_in":13107,"tokens_out":3352,"would_cite":false,"duration_ms":31227,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a hybrid quantum-classical network maps seven bridge sensors to 1,017 nodal displacements with test MSE 3.16e-11, beating classical baselines and running roughly 2,000 times faster than one finite element solve.","keywords":["quantum machine learning","digital twin","inverse finite element analysis","structural health monitoring","symmetric positive definite matrices","parameterized quantum circuit","surrogate modeling","bridge monitoring"],"falsifier":"Recompute the test-set target standard deviation from the reported RMSE and NRMSE(std) for the best QMLP row; the values imply a standard deviation near $7.7 \\times 10^{-4}$, which would yield $R^2$ near 0.99995 rather than the reported 0.9856. Re-running all models on one fixed test split with identical target scaling and reporting raw prediction files would decide whether the quantum advantage is real.","tokens_in":12072,"feed_emoji":"🌉","tokens_out":8473,"duration_ms":95782,"temperature":0.7,"pith_summary":"The paper tries to establish that a hybrid quantum-classical neural network can act as a real-time surrogate for inverse finite element analysis in digital twins. The specific problem is mapping seven tilt-sensor readings from a bridge to 1,017 nodal displacements, where classical finite element inversion is slow and ill-posed. The proposed pipeline embeds sensor data into symmetric positive definite matrices via polynomial features, converts that into a quantum state through Hilbert-Schmidt mapping, processes it with a parameterized quantum circuit, and finishes with a classical neural network. The paper reports that this QMLP reaches MSE $3.16 \\times 10^{-11}$ on the bridge test set, far below the classical MLP baseline near $10^{-3}$, and that inference takes milliseconds instead of roughly 30 seconds for one FE step. If the reported numbers hold, this would let digital twins update full structural displacement fields in near real time from sparse sensors.","feed_headline":"Quantum-classical net maps 7 bridge sensors to 1,017 displacements","feed_subtitle":"It predicts 1,017 nodal displacements in milliseconds, about 2,000x faster than a finite element solve, if the reported numbers hold","key_machinery":"The load-bearing object is the symmetric positive definite (SPD) matrix built from a polynomial expansion of the seven sensor readings. The outer product of the expanded feature vector with itself, plus a small diagonal regularization, yields a positive definite matrix; taking its matrix square root and normalizing by the trace gives a valid density matrix. Hilbert-Schmidt vectorization then flattens that density matrix into a quantum state, and a parameterized quantum circuit with angle embedding and entangling layers transforms the state. Pauli expectation values from the circuit are passed to a classical network that outputs the 1,017 displacements. The pipeline's claim is that the SPD geometry and Hilbert-Schmidt embedding preserve nonlinear structure that a plain MLP misses.","core_discovery":"On the paper's own terms, the discovery is that a hybrid quantum-classical multilayer perceptron can solve the inverse finite element problem end-to-end: seven tilt-meter readings from a bridge map to 1,017 nodal displacements with reported MSE $3.16 \\times 10^{-11}$, beating purely classical baselines and running roughly 2,000 times faster than a single FE solve. The quantum part is not a standalone solver; it is a feature extractor whose outputs feed a classical regressor, and the paper attributes its accuracy to polynomial-plus-SPD embedding and Hilbert-Schmidt quantum encoding rather than to quantum speedup.","pith_inferences":["Beyond the paper: if the pipeline transfers, the same SPD-polynomial plus variational-circuit stack could be retrained for pipelines, offshore platforms, and other sensor-to-field inverse problems, because the dimensional mismatch has the same shape.","Beyond the paper: the reported 2,000x speedup is measured against a forward FE solve, not against an inverse solver; a deployment comparison would need to benchmark against the fastest classical inverse-surrogate alternative.","Beyond the paper: an independent re-run with a published fixed split, identical target scaling, and raw prediction files would settle whether the headline MSE is reproducible; the paper's printed metrics are not internally consistent."],"forward_implications":["Inverse FE mapping from sparse sensors can run in milliseconds, so digital-twin updates can happen near real time rather than after a 30-second FE solve.","At the reported RMSE around $5.62 \\times 10^{-6}$, predicted displacements across the 339 nodes would resolve sub-millimeter structural motion.","The full hybrid pipeline, at roughly 37.5 ms per sample, is about 2,000 times faster than one FE analysis step on the same bridge.","The same SPD-polynomial plus variational-circuit stack can be retrained for other large-scale structures with the same low-to-high dimensional sensor-to-field mapping."],"supporting_citations":[{"why":"Supplies the validated FE model and the 10,000-case sensor-to-displacement dataset used for training and testing.","marker":"11"},{"why":"Provides the SPD-on-Riemannian-manifold encoding that this paper extends with polynomial features and Hilbert-Schmidt embedding.","marker":"31"},{"why":"Gives the foundational quantum neural network formulation on which the parameterized circuit layer is built.","marker":"34"},{"why":"Sets out the design requirements and challenges for quantum neural networks that shape the PQC architecture.","marker":"35"},{"why":"Provides the spectral decomposition used to form the SPD/density matrix and its complexity estimate.","marker":"40"},{"why":"Supplies the gate-model operation counts used to estimate the quantum layer's per-sample inference cost.","marker":"41"}],"fun_headline_variants":["Hybrid quantum net runs bridge FE inverse 2000x faster","Quantum-classical surrogate solves bridge inverse FE in ms","QMLP beats classical baselines on bridge inverse mapping","Quantum feature extractor speeds digital twin updates 2000x","7 sensors to 1,017 displacements: quantum net does it 2000x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim collapses if the rows of the comparison table were not all evaluated on the same test set with the same target scaling; in the paper's best QMLP row the printed RMSE, NRMSE(std), and $R^2$ are not algebraically consistent, so that assumption is load-bearing and unverified.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid quantum net runs bridge FE inverse 2000x faster","Quantum-classical surrogate solves bridge inverse FE in ms","QMLP beats classical baselines on bridge inverse mapping","Quantum feature extractor speeds digital twin updates 2000x","7 sensors to 1,017 displacements: quantum net does it 2000x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001086,"raw_usage":{"total_tokens":4535,"prompt_tokens":934,"completion_tokens":3601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":3511}},"tokens_in":550,"tokens_out":3601,"duration_ms":27344,"temperature":1.0,"reasoning_tokens":3511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:45:59.385012+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the test-set target standard deviation from the reported RMSE and NRMSE(std) for the best QMLP row; the values imply a standard deviation near $7.7 \\times 10^{-4}$, which would yield $R^2$ near 0.99995 rather than the reported 0.9856. Re-running all models on one fixed test split with identical target scaling and reporting raw prediction files would decide whether the quantum advantage is real.","supporting_citations":[{"cited_title":"�� ���Innovative digital twin with artificial neural networks for real-time monitoring of structural response: A port structure case study","cited_arxiv_id":null,"evidence_quote":"Supplies the validated FE model and the 10,000-case sensor-to-displacement dataset used for training and testing."},{"cited_title":"& Mahmoodian, M","cited_arxiv_id":null,"evidence_quote":"Provides the SPD-on-Riemannian-manifold encoding that this paper extends with polynomial features and Hilbert-Schmidt embedding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the foundational quantum neural network formulation on which the parameterized circuit layer is built."},{"cited_title":"& Petruccione, F","cited_arxiv_id":null,"evidence_quote":"Sets out the design requirements and challenges for quantum neural networks that shape the PQC architecture."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the spectral decomposition used to form the SPD/density matrix and its complexity estimate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gate-model operation counts used to estimate the quantum layer's per-sample inference cost."}],"review_version":1}