{"id":"5d3a955c-ef88-4913-b49c-47f873bd4fea","arxiv_id":"2506.22845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"On one wind turbine dataset, simulated 4-qubit quantum neural networks with a Z feature map reach R2 around 0.94 and RMSE slightly below k-nearest neighbors, with training time scaling roughly linearly with dataset size.","lead":"This paper benchmarks six small quantum neural network circuits against three classical machine learning models for predicting wind turbine power output. The quantum circuits match or slightly beat the simple classical baselines on this dataset, and simulation time grows roughly linearly with dataset size.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"QNN-3's claimed 4.5% holdout RMSE advantage over kNN is within the paper's own 5-fold CV noise, so the headline 'marginally better' is not supported without repeated-seed significance testing.","rationale":"The reader's weakest assumption about scaler leakage is legitimate and should be checked, so I agree with it as a reproducibility and preprocessing risk. However, I see an even more direct, internally visible soft spot: the headline 4.5% QNN-3 advantage is smaller than the fold-to-fold variation in the paper's own cross-validation tables, and the holdout comparison is a single split with no significance testing. The paper does not show that QNN-3 is reliably better than kNN; it shows that QNN-3 is competitive with kNN within noise. This does not make the central claim false: across cross-validation and holdout tests, the QNN configurations do reach R2 values around 0.89–0.94 and RMSE values broadly comparable to the classical baselines, and the empirical linear scaling of simulation time with dataset size is a reasonable observation from Table 13 and Figure 5. The appropriate verdict remains conditional, as the reader concluded: the claim should be accepted only with code/data release, an explicit statement of scaler fitting, and a repeated-seed paired comparison. My read sharpens the condition but does not move the verdict; hence UNCHANGED.","tokens_in":23808,"tokens_out":7368,"duration_ms":80580,"concrete_test":"Re-run the dataset-size-4000 experiment with the min-max scaler fitted on the 3200-sample training split only, and with inverse transformation of predictions using training-target min/max only, for at least 20 random 80/20 splits and 20 QNN initializations. For each split, train QNN-3 and kNN on the same training data and record the paired holdout RMSE difference (RMSE_QNN3 - RMSE_kNN). Report the distribution of paired differences and its 95% confidence interval. If the confidence interval excludes 0 in QNN-3's favor and the median improvement is at least the claimed 4.5%, the headline claim is supported; if the interval crosses 0, the correct conclusion is parity within noise and the abstract's 'marginally better' wording should be removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 4.4, Table 19) is that QNN-3 beats kNN on the 800-sample holdout set with RMSE 174.67 versus 182.99 kW, a 4.5% error reduction. This gap is smaller than the variability reported by the paper's own 5-fold cross-validation on the same training data: on the 3200-sample training set, QNN-3 has mean CV RMSE 207.64 ± 27.71 kW (Table 11) while kNN has mean CV RMSE 204.18 ± 12.68 kW (Table 18). That is, on the cross-validation folds kNN is actually slightly better on average, and the individual fold RMSE ranges overlap heavily (QNN-3: 181.78–254.09; kNN: 186.93–220.88). Table 19 is a single 80/20 split with a single training run and reports no seed, no repeated splits, and no paired significance analysis, so the 8.32 kW gap cannot be distinguished from split-to-split and initialization noise. This is load-bearing because the abstract's 'marginally better' claim rests entirely on that gap; the paper's own cross-validation evidence supports at most 'competitive within noise.' Separately, Section 2 states only that data were divided into training and test sets and then min-max scaled, without saying whether the scaler was fit on the training split alone; if it was fit on combined train+test data, the holdout numbers would be optimistically biased. The concrete test below addresses both issues.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an empirical benchmark of six quantum neural network (QNN) configurations—all using the Z Feature Map with different ansatz entanglement strategies—against three classical regressors (kNN, DTR, LR) for wind turbine power output prediction. Experiments are run on random subsets of 1000 to 4000 samples, with 5-fold cross-validation on training portions and a single 80/20 holdout evaluation. The central claims are that QNNs achieve predictive performance competitive with, and in some cases marginally better than, the classical baselines (with QNN-3 reaching R²=0.94 and RMSE=174.67 kW on the 800-sample holdout versus kNN's 182.99 kW), and that QNN simulation time scales linearly, O(n), with dataset size.","tokens_in":24144,"tokens_out":3882,"duration_ms":41591,"significance":"If fully supported, the study would provide useful empirical guidance for applying QNNs to small tabular regression tasks in the energy domain. The manuscript has notable strengths: fold-level cross-validation tables are reported for every configuration and dataset size, circuit diagrams and gate counts are documented for reproducibility, and the authors explicitly acknowledge limitations such as the absence of hyperparameter tuning and unphysical negative predictions. The comparison across dataset sizes and the gate-count versus runtime analysis are also potentially informative. However, the headline quantitative claims—particularly the 'marginally better' comparison and the O(n) complexity statement—are not adequately supported by the evidence as presented, so the paper is currently most defensible as a demonstration that QNNs are broadly competitive within noise, rather than that they surpass classical methods.","major_comments":[{"comment":"The claim that QNNs are 'in some cases marginally better' than classical models rests on a single 80/20 holdout evaluation with no repeated splits, no multiple random seeds, and no significance test. In Table 19, QNN-3 achieves RMSE 174.67 kW versus kNN's 182.99 kW on the 800-sample holdout, a gap of 8.32 kW (4.5%). This gap is smaller than the variability in the paper's own 5-fold cross-validation on the corresponding 3200-sample training set: QNN-3 has mean CV RMSE 207.64 ± 27.71 kW (Table 11) and kNN has 204.18 ± 12.68 kW (Table 18), with heavily overlapping fold ranges. Thus the holdout difference cannot be distinguished from split-to-split or initialization noise, and the abstract's 'marginally better' phrasing and §4.5's stronger language ('significant advantage') are not supported. The authors should report repeated holdout evaluations (e.g., multiple random splits or repeated k-fold) with paired tests such as a Wilcoxon signed-rank test, and adjust the conclusions to 'competitive within noise' unless such tests justify a stronger claim.","section":"§4.4, Table 19 vs. Tables 11 and 18"},{"comment":"The preprocessing description states that data were 'divided into training (80%) and test (20%) sets, and then used a min-max scaler (Equation (1)) to scale all values' but does not state whether the scaler's x_min and x_max were computed on the training split alone or on the combined training and test sets. If the scaler is fit using test-set statistics, information from the holdout set enters the training pipeline, which would optimistically bias every holdout number in Table 19, including the headline QNN-3 versus kNN comparison. The authors must clarify that scaling was fit only on the training data, or, if it was not, rerun the experiments with correct train-only scaling and report the scaler-fitting code or explicit procedure.","section":"Section 2, Data Preprocessing"},{"comment":"The claim of 'time complexity O(n)' is based on a linear regression fitted to only four dataset-size points (800, 1600, 2400, 3200) with no reported goodness-of-fit, residuals, confidence intervals, or comparison against alternative scaling functions such as quadratic or log-linear fits. With four points, the linear trend may simply be an artifact of the chosen sizes and the specific simulator. The claim as stated in §4.2.1 and repeated in §5 ('computational complexity of O(n)') is too strong. The authors should provide more granular runtime measurements (ideally multiple independent runs per dataset size), report fit statistics and residual analysis, and rephrase the conclusion as an empirical observation of approximately linear growth within the tested range rather than an asymptotic complexity statement.","section":"§4.2.1, Equation (13) and Table 14"}],"minor_comments":[{"comment":"There is a duplicated word in the sentence 'are used used for finding optimal parameters'; this should be corrected.","section":"Section 3.2.4"},{"comment":"The stability score SC in Equation (12) is defined as the sum of min-max normalized standard deviation, maximum spike after iteration 10, and final loss, but the normalization range and the direction of the normalization (whether lower raw values always give lower normalized values) are not defined precisely. Please specify the normalization procedure or give explicit formulas.","section":"Equation (12), Table 12"},{"comment":"The phrase 'in 1' in the sentence 'there are four options, as shown in 1 which are:' is incomplete; it should reference Figure 1.","section":"Section 1, End of Introduction"},{"comment":"The selection of the 10th iteration as the threshold for 'convergence' is arbitrary; please justify this choice or include a sensitivity check, since the stability score depends on it.","section":"Section 4.1.3"},{"comment":"The paper reports that linear regression and decision tree regressors use 'default parameters' but does not specify the scikit-learn version or the exact defaults in use; adding version numbers would improve reproducibility.","section":"Section 3.4"},{"comment":"The manuscript does not specify the number of shots, the measurement basis, or how the expectation value is mapped to a continuous power prediction in the regression task. These details are essential for reproducing the QNN results and should be added.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a straightforward empirical benchmark with limited novelty relative to the authors' prior work, but the comparison methodology and transparency about fold-level results make it potentially suitable for an applied QML venue. The main concerns are statistical overclaiming and potential scaling leakage; both are fixable within the manuscript's scope. I would not recommend rejection, as the qualitative conclusion of competitiveness is defensible once the statistical claims are softened and the preprocessing is clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful, honest benchmark, but its one overreach is the word \"marginally better\" in the abstract. The 4.5% holdout RMSE gap between QNN-3 and kNN (174.67 vs 182.99 kW, Table 19) comes from a single 80/20 split with no repeated seeds and no significance test. The paper's own 5-fold CV on the 3200-sample training set shows kNN at 204.18 ± 12.68 and QNN-3 at 207.64 ± 27.71, with fold ranges that overlap heavily. So the supportable conclusion is \"competitive within noise,\" not \"better.\" The stress-test note is right about this, and it is load-bearing because the abstract makes the marginal-gain claim.\n\nWhat is actually new: the dataset-size sweep (800/1600/2400/3200), the time-scaling comparison, and the stability ranking. These extend the authors' earlier paper on the same data rather than redo it. The reporting is unusually transparent in places: full CV tables, gate counts, circuit diagrams, acknowledgment of negative predictions, and admission that no hyperparameter tuning was done on either side. That earns credit.\n\nSoft spots, in order:\n1. Preprocessing (Section 2) says data were split then min-max scaled, but never says the scaler was fit on training only. If it was fit on combined train+test, the holdout numbers are optimistically biased. Easy to fix with code or a sentence.\n2. The headline comparison is one split. Need repeated splits and paired tests, or at least reported seeds.\n3. The O(n) time-complexity statement is a linear fit to four points on one simulator, not a complexity result. Fine as an empirical observation, but should be labeled as such.\n4. Single dataset, four features, simple classical baselines (default kNN/DT/LR). That limits scope, though the authors mostly say so.\n\nWho this is for: practitioners choosing ansatz structures for small tabular QNN benchmarks, and readers in energy-domain QML. It deserves a serious referee because the benchmark is clean enough to verify and the tables are internally consistent. I would send it to review with a request for code, scaler handling, seeds, and significance testing. It is not a quantum-advantage paper and should not be treated as one.","headline":"Useful and honest QNN benchmark, but the 'marginally better' claim rests on a single split and is not supported by the paper's own cross-validation.","tokens_in":24640,"tokens_out":2699,"would_cite":false,"duration_ms":31518,"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":"This paper claims six QNN configurations match or slightly beat classical regressors on wind-turbine power prediction, with QNN-3 reaching R²=0.94 and RMSE 174.67 kW versus kNN's 182.99 kW.","keywords":["quantum neural network","variational quantum circuit","wind power forecasting","Z feature map","ansatz entanglement","quantum machine learning","simulation time scalability","regression benchmark"],"falsifier":"Re-run the entire benchmark, for all six QNNs and the three classical models at the 4000-sample size, with the min-max scaler fitted on the 80% training split and then applied unchanged to the hold-out set; if QNN-3's RMSE no longer stays below kNN's 182.99 kW, the central comparative claim loses its evidence. A second check is to repeat the data-splitting and scaling procedure over several random seeds to see whether QNN-3's 1600-sample cross-validation peak reflects a stable property or the single split used in the paper.","tokens_in":23629,"feed_emoji":"🌬️","tokens_out":9661,"duration_ms":91085,"temperature":0.7,"pith_summary":"This paper tries to establish that quantum neural networks (QNNs), built as variational quantum circuits trained by a classical optimizer, are a competitive alternative to classical regressors for wind-turbine power prediction. On a real dataset of 4464 ten-minute turbine readings, six QNN configurations built on the Z Feature Map with different entanglement patterns are benchmarked against k-nearest neighbours, decision trees, and linear regression at four dataset sizes. The headline result is that QNN-3, the circular-entanglement configuration, reaches $R^2=0.94$ and $\\mathrm{RMSE}=174.67$ kW on the 800-sample hold-out, versus kNN's 182.99 kW, a 4.5% error reduction. The paper also claims simulation time grows roughly linearly with dataset size and is controlled mainly by total gate count. If these claims hold, QNNs offer a viable, cheap-to-simulate option for small tabular forecasting problems in the energy domain.","feed_headline":"Quantum circuit beats best classical wind forecast by 4.5%","feed_subtitle":"The best QNN reaches R²=0.94 with RMSE 174.67 kW on 800 holdout samples, versus kNN's 182.99 kW.","key_machinery":"The central object is the Z Feature Map, a data-encoding circuit that maps each classical feature into a single-qubit phase via Hadamard gates and $Z$-axis rotations (two repetitions), followed by a RealAmplitudes ansatz whose $Y$-rotations and $\\mathrm{CNOT}$ entangling gates are arranged into six entanglement strategies: Full, Linear, Circular, Shifted-Circular-Alternating, Reverse Linear, and Pairwise. Keeping the encoding fixed and varying only the ansatz isolates the effect of circuit geometry and gate count on accuracy and runtime. A classical optimizer (L-BFGS-B) trains the circuit parameters, and the total gate count for each configuration ranges from 34 to 40, which the paper uses as the main explanatory variable for simulation time.","core_discovery":"The paper's central claim is experimental: given a fixed data-encoding circuit (the Z Feature Map) and six ansatz layouts that differ only in entanglement strategy, QNNs can match or slightly outperform classical benchmarks on wind-power regression. The strongest specific result is QNN-3 on the largest hold-out set, with $R^2=0.94$ and $\\mathrm{RMSE}=174.67$ kW against kNN's 0.94 and 182.99 kW; the paper also reports that quantum models gave the lowest or equal-lowest RMSE across all tested hold-out sizes. It further claims that cross-validated QNN performance peaks at 1600 training samples and then plateaus or declines, that all six configurations converge within roughly 10–15 L-BFGS-B iterations, that simulation time scales as $O(n)$ with dataset size, and that total gate count, more than gate arrangement, determines runtime. The paper acknowledges that QNNs sometimes output negative power values, which are physically impossible, and that no hyperparameter tuning was performed for either quantum or classical models.","pith_inferences":["Editorial inference: the reported 4.5% advantage rests on a preprocessing detail the paper never states explicitly; if the min-max scaler was fit on the full dataset rather than the training split only, the hold-out numbers would be optimistically biased and the gap could shrink or disappear.","Editorial inference: the $O(n)$ time scaling is a classical-simulation result tied to gate count; on real quantum hardware the bottleneck shifts to measurement shots, noise, and qubit connectivity, so the scalability conclusion should not be assumed to transfer.","Editorial inference: the performance peak at 1600 samples may be a property of these specific shallow four-qubit circuits; testing deeper ansätze or more expressive feature maps could shift or erase the peak, connecting this measurement to expressibility and trainability questions.","Editorial inference: because the classical baselines were run with default parameters and hyperparameter search was explicitly left out for all models, a fair tuned comparison (including stronger regressors such as gradient-boosted trees) is the natural next test of whether the quantum advantage survives."],"forward_implications":["On this wind-turbine regression task, QNN configurations can match or slightly beat kNN, decision trees, and linear regression, so they are a plausible alternative for small tabular forecasting problems.","QNN training time on classical simulators scales approximately linearly with dataset size, and designs with fewer total gates simulate fastest, independent of entanglement geometry.","Adding training data beyond about 1600 samples did not improve QNN accuracy and sometimes worsened it, suggesting an optimal data-to-model-complexity ratio for these shallow circuits.","All six QNNs converge within about 10–15 iterations with the chosen optimizer, and larger datasets reduce training-loss variance.","Because QNNs can output physically impossible negative power predictions, deployment would require added constraints such as a ReLU output or a penalty in the loss function."],"supporting_citations":[{"why":"Supplies the earlier comparison finding that the Z Feature Map outperforms the ZZ Feature Map on this dataset and that QNNs can compete with kNN, decision trees, and linear regression; the present study extends that benchmark.","marker":"[40]"},{"why":"Provides the wind-turbine dataset and the three classical baseline models (kNN, DTR, LR) with their parameter settings that the QNNs are compared against.","marker":"[41]"},{"why":"Supplies the quantum feature-map formalism that defines the Z Feature Map used for data encoding in all six configurations.","marker":"[47]"},{"why":"Provides the quantum circuit construction and simulation library in which all QNN configurations were implemented and timed.","marker":"[48]"},{"why":"Provides the QNN training loop (the quantum-classical hybrid component) used for all six configurations in the experiments.","marker":"[53]"}],"fun_headline_variants":["Quantum neural nets match classical wind forecasting","QNNs rival classical models in wind power prediction","Wind forecast: QNNs equal or beat classical benchmarks","Quantum circuit trims wind forecast error by 4.5%","QNNs show competitive wind energy forecasting vs classical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the min-max rescaling (which maps each feature into $[0,1]$) was fitted using only the 80% training portion; the paper says the data were split and then scaled but never says the scaler was fit on the training split, and if test statistics leaked into the scaling, the reported hold-out $R^2$ and RMSE values would be optimistically biased.","fun_headline_variants_meta":{"raw":{"variants":["Quantum neural nets match classical wind forecasting","QNNs rival classical models in wind power prediction","Wind forecast: QNNs equal or beat classical benchmarks","Quantum circuit trims wind forecast error by 4.5%","QNNs show competitive wind energy forecasting vs classical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1879,"prompt_tokens":998,"completion_tokens":881,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":804}},"tokens_in":614,"tokens_out":881,"duration_ms":9215,"temperature":1.0,"reasoning_tokens":804,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:57:33.764056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the entire benchmark, for all six QNNs and the three classical models at the 4000-sample size, with the min-max scaler fitted on the 80% training split and then applied unchanged to the hold-out set; if QNN-3's RMSE no longer stays below kNN's 182.99 kW, the central comparative claim loses its evidence. A second check is to repeat the data-splitting and scaling procedure over several random seeds to see whether QNN-3's 1600-sample cross-validation peak reflects a stable property or the single split used in the paper.","supporting_citations":[{"cited_title":"Comparative Analysis of QNN Architectures for Wind Power Prediction: Feature Maps and Ansatz Configurations","cited_arxiv_id":"2506.14795","evidence_quote":"Supplies the earlier comparison finding that the Z Feature Map outperforms the ZZ Feature Map on this dataset and that QNNs can compete with kNN, decision trees, and linear regression; the present study extends that benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the wind-turbine dataset and the three classical baseline models (kNN, DTR, LR) with their parameter settings that the QNNs are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quantum feature-map formalism that defines the Z Feature Map used for data encoding in all six configurations."}],"review_version":1}