{"id":"f60f9fac-fce4-42b6-a32f-521605ee15d8","arxiv_id":"2602.00525","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Classical and quantum classifiers label energy points from the TDDFT absorption spectra of pure vs Er-doped CaF2, with classical SVM at 0.983 test accuracy, but the benchmark uses only two spectra and is therefore not a meaningful generalization test.","lead":"This paper computes optical absorption spectra for pure and erbium-doped CaF2 clusters with DFT and TDDFT, then uses those spectra as features to train classical and quantum classifiers. A generalist might read it as a worked quantum-ML materials benchmark, but the dataset contains only two computed spectra, so the reported accuracies do not support the broad 'robust feature space' claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark claim is invalid because train/test points are correlated samples from two smooth spectra; reported accuracies reflect interpolation, not generalization.","rationale":"The reader identified the dependence of the 1,589 energy-resolved points and the resulting lack of statistical power as the weakest assumption; my independent stress-test reaches the same conclusion: the dataset is two smooth deterministic absorption curves sampled densely, so random point-wise splits create train/test interpolation rather than generalization. The abstract's claim of a 'robust, physically grounded feature space for benchmarking near-term quantum learning models' is exactly what fails if the evaluation cannot support generalization. I see no other load-bearing flaw of comparable weight: the TDDFT workflow is plausible, the ML implementations appear competent, and the hardware section is appropriately cautious. The paper's own text (§II.B, §III.E) implicitly acknowledges limitations but does not connect them to the central benchmarking claim. I recommend REJECT (not CONDITIONAL) because the flaw is in the core evaluation design: a benchmark that cannot distinguish interpolation from generalization has no valid accuracy numbers to report, so the stated contribution cannot stand as written. A repair is possible (multiple cluster sizes, leave-one-system-out with more systems, or explicit reframing as a two-curve study), but requiring those changes is a rejection of the current manuscript, not a minor revision.","tokens_in":19479,"tokens_out":2784,"duration_ms":25435,"concrete_test":"Retrain the classical SVM and the QNN on the existing 1,589-point-per-system dataset but test on held-out energy windows (e.g., train on E in [0,5] eV ∪ [7,10] eV, test on E in [5,7] eV) for both systems, or equivalently use leave-one-system-out by training on one material and testing on the other. If accuracy drops toward chance on held-out windows/systems, the point-wise split was the source of the reported performance and the benchmark claim fails. If accuracy remains high, the descriptors do generalize across spectral regions and the reader's concern is mitigated.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that the TDDFT-derived descriptors form a \"robust, physically grounded feature space for benchmarking near-term quantum learning models\" rests entirely on the reported accuracies (classical SVM 0.983 AUC 0.999; QSVM 0.851/0.817; hybrid QNN 0.93). But the dataset is constructed by sampling 1,589 energy points along two smooth, deterministic absorption curves (Figure 8; Table I), one per material. Under the random point-wise split used in §III.B–C, test points sit at energies immediately adjacent to training points on the same curve; because the descriptors (E, κ, α) are continuous functions of energy with correlation length much larger than the sampling spacing, the classifier is effectively interpolating the two known spectra. The reported near-perfect classical accuracy and the QML accuracies therefore do not measure generalization to unseen materials, dopant concentrations, cluster sizes, or noise levels—they measure separation of two curves. The paper itself concedes the limited scope at §III.E (small hardware slice \"cannot be considered definitive\") and §II.B (a \"set instead of an unordered sequence\"), but never treats the two-spectrum structure as the evaluation unit. The bootstrap confidence intervals in §III.C–D (e.g., [0.965, 0.997]) resample from the same dependent points, so they do not address between-spectrum variability. The load-bearing weakness is not the physics (TDDFT workflow is plausible) but the statistical independence assumption: with only two systems, any point-wise split has no power to establish a transferable feature space. The claim could be repaired by evaluating at the system level (leave-one-system-out would be trivial with N=2 and cannot support the claim), by generating multiple cluster sizes/dopant concentrations, or by explicitly reframing the result as a two-curve interpolation exercise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper constructs finite clusters of pristine CaF2 (Ca8F16) and Er-doped CaF2 (Ca7ErF16), computes their optical spectra with LR-TDDFT, and derives energy-resolved descriptors (transition energy E, extinction coefficient κ, absorption coefficient α) at 1,589 points per system. These descriptors are fed into a classical RBF SVM, quantum SVMs (statevector, noisy QASM, IBM hardware), and a hybrid quantum neural network. The reported accuracies are 0.983 (classical SVM), 0.851/0.817 (QSVM simulators), 0.733 (hardware QSVM on a 15-sample test slice), and 0.93 (hybrid QNN). The paper claims these results establish a robust, physically grounded feature space for benchmarking near-term quantum learning models against classical baselines.","tokens_in":19729,"tokens_out":1865,"duration_ms":23678,"significance":"If the benchmark were valid, the work would provide a useful template for combining first-principles optical descriptors with quantum classifiers, and the comparison of classical, simulator, and hardware results would inform NISQ-era materials informatics. The TDDFT workflow, including cluster construction, Casida calculations, and Gaussian broadening, is described in sufficient detail to be reproduced, and the authors provide code and data. However, the central statistical claim is undermined by the data construction: the 1,589 samples per class are not independent measurements but densely sampled points along two smooth, deterministic absorption curves. Consequently, the high classification accuracies largely reflect interpolation of two known spectra rather than generalization to new materials, dopant concentrations, or noise levels. This invalidates the abstract's claim that the feature space is 'robust' and that the benchmarks are meaningful. The paper's internal caveats (§III.E: the hardware slice 'cannot be considered definitive'; §II.B: the descriptor set is 'a set instead of an unordered sequence') do not address this more fundamental limitation.","major_comments":[{"comment":"The dataset consists of 1,589 energy-resolved points from each of only two smooth absorption spectra, one per material. Under the random point-wise train/test split used in §III.B–C, test points lie at energies immediately adjacent to training points on the same curve. Because the descriptors E, κ, α are continuous functions of energy with correlation length set by the Gaussian broadening σ=0.1–0.2 eV (Eq. 4) and the sampling spacing is much finer, the classifiers interpolate two known spectra. The reported test accuracy (0.983), AUC (0.999), and QML accuracies therefore do not measure generalization to unseen materials or configurations; they measure separability of two curves. The manuscript nowhere treats the two spectra as the evaluation unit. This is the load-bearing issue for the central claim that the descriptors form a 'robust feature space.'","section":"§II.A, §III.B, Table I, Figure 8"},{"comment":"The bootstrap confidence intervals (e.g., classical SVM mean 0.983, 95% CI [0.965, 0.997]; QSVM SV [0.810, 0.893]) resample from the same dependent point dataset. Since the points are autocorrelated along each spectrum, the bootstrap only reflects sampling variability among correlated points on two fixed curves, not between-spectrum variability. It provides no evidence about how the models would perform on other dopant concentrations, cluster sizes, or disorder realizations. This does not rescue the generalization claim.","section":"§III.C–D, bootstrap confidence intervals"},{"comment":"The conclusion that dopant-induced optical fingerprints form a 'robust, physically grounded feature space for benchmarking near-term quantum learning models against strong classical baselines' is not supported by the experimental design. With only two systems, any sufficiently flexible classifier can achieve near-perfect separation on a fine energy grid. The physical descriptors may indeed be informative, but the benchmark as constructed cannot validate that claim. Demonstrating robustness would require multiple doped/undoped instances (different concentrations, cluster sizes, or host variants) with evaluation at the spectrum level, not at the level of individual energy points.","section":"Abstract, §III.C, §III.F"}],"minor_comments":[{"comment":"The hardware QSVM uses only 30 training and 15 test samples; the authors correctly note that these results 'cannot be considered definitive,' but the small sample size should be more prominently featured in the abstract and conclusion, where the hardware accuracy is reported without qualification.","section":"§III.E"},{"comment":"Feature selection via linear SVM weights is performed on the data before describing the train/test split. If the weight-based ranking uses all 3,178 points, the selected features (α, κ, E) may incorporate information from the test set. The authors should clarify whether feature selection was nested inside the training folds or performed on the full dataset.","section":"§II.B and Figure 9"},{"comment":"Some notation is introduced without explicit definition (e.g., 'N_{occ}' in Eq. 1, 'E_min' in Eq. 21). The paper would benefit from a table summarizing all hyperparameters and dataset sizes for each model, including the QNN's effective training sample count of 2,304 points and the relationship between this number and the 1,589-point per-system dataset.","section":"General"},{"comment":"The discussion compares QSVM and QNN accuracies across heterogeneous datasets (MNIST, hyperspectral images, polymer band gaps, ADME-Tox). These comparisons are not quantitatively meaningful without controlling for task difficulty, data size, and evaluation protocol. The text should clearly state that cross-dataset accuracy comparisons are illustrative only.","section":"§IV and Table II"}],"recommendation":"reject","confidential_remarks":"The core issue is not the TDDFT calculations or the ML implementations, which appear plausible, but the statistical design: two smooth spectra sampled densely yield an effectively two-sample classification problem. No amount of resampling or reanalysis of the existing data can convert this into a valid benchmark of generalization. The central claim of the paper therefore rests on an unfixable design flaw within the manuscript's current scope. A future study with many independent material configurations, evaluated at the spectrum level, could re-purpose the workflow, but that is a new dataset and a new paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nTwo things to know upfront. The paper applies a standard TDDFT-to-classical/QML pipeline to distinguish pristine CaF2 from Er-doped CaF2, and the electronic-structure side is plausible. But the central benchmark is broken: they take 1,589 energy-resolved points from each of exactly two smooth absorption spectra and split them randomly into train/test. Because the descriptors are continuous functions of energy and the sampling is dense, test points sit immediately next to training points on the same curve. The classifiers are interpolating the two known spectra, which is why the classical SVM gets 0.983 accuracy and 0.999 AUC. The two curves are already visibly different in Figure 8; the high accuracies carry no information about generalization to new materials, dopant concentrations, or noise levels.\n\nWhat is actually new: not much. The combination of TDDFT descriptors with QSVM/QNN is a routine extension of the QML benchmark literature the authors themselves cite. The circuits are stock. That said, the paper earns credit for a clear workflow, concrete parameters, code/data availability, and for explicitly saying QML is not expected to outperform classical. They also own up to the small hardware test slice in Section III.E. That honesty makes the bigger omission more frustrating: they never treat the number of physical systems as the sample size.\n\nThe load-bearing claim—'dopant-induced optical fingerprints form a robust, physically grounded feature space for benchmarking'—does not follow from the evidence. With N=2 systems, no point-wise split can establish a benchmark. Leave-one-system-out is trivial and unhelpful. The bootstrap confidence intervals just resample the same dependent points.\n\nMinor soft spots: the feature-selection step uses a linear SVM to pick features for an SVM, which is a bit circular but not fatal. The QNN accuracy of 0.93 is on the same two curves, so it doesn't tell us anything about quantum advantage or representational power.\n\nThis is a decent worked example of a pipeline, not a scientific result. The authors could repair it by generating multiple cluster sizes/dopant concentrations and evaluating at the system level, or by reframing the paper as an interpolation exercise. I would not send this version to peer review; I'd desk reject with an invitation to resubmit after fixing the evaluation design. It's worth citing as a cautionary example of data leakage in QML benchmarks, but not as a validated benchmark.","headline":"Plausible TDDFT pipeline, invalid benchmark: 1,589 points from two smooth spectra are treated as independent, so the accuracies just interpolate two visibly different curves.","tokens_in":20398,"tokens_out":4252,"would_cite":false,"duration_ms":49136,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Er dopant fingerprints split CaF2 spectra at 98% accuracy","keywords":["Er-doped CaF2","TDDFT absorption spectra","optical descriptors","quantum machine learning","quantum support vector machine","quantum neural network","materials classification","spectral fingerprints"],"falsifier":"Run the same classifiers on a dataset where the energy grid is coarsely sampled (so neighboring training and test points are not adjacent) or split by contiguous energy blocks; if accuracy collapses or the classical near-perfect separation vanishes, the benchmark is measuring interpolation of two smooth spectra rather than learned discrimination.","tokens_in":19310,"feed_emoji":"⚛️","tokens_out":5634,"duration_ms":65496,"temperature":0.7,"pith_summary":"The paper tries to establish that the optical fingerprints created by putting a single erbium atom in a calcium fluoride cluster—computed from first principles, not fitted to experiment—give a reliable, physically interpretable feature space for machine-learning classification, and that this feature space can serve as a meaningful benchmark for comparing near-term quantum classifiers against a classical baseline. Using time-dependent density-functional theory on a 24-atom cluster, the authors compute absorption spectra for pristine and Er-doped CaF2, extract three descriptors, and show a classical SVM separates the two classes almost perfectly, with 98.3% test accuracy. Quantum support-vector machines reach 85.1% on an ideal simulator, 81.7% under simulated noise, and 73.3% on real quantum hardware; a trainable hybrid quantum network reaches 93%. The point is not that quantum wins—it does not—but that first-principles spectral descriptors give a controlled, physically grounded setting in which to measure how much near-term quantum models lose to strong classical methods.","feed_headline":"Er dopant fingerprints split CaF2 spectra at 98% accuracy","feed_subtitle":"A classical model wins; quantum classifiers hit 85% in simulation and 73% on real hardware.","key_machinery":"The carrying object is the Gaussian-broadened optical absorption spectrum computed by linear-response time-dependent density-functional theory for a 2x2x2 fluorite cluster with and without a substitutional erbium dopant. The mechanism that creates class separability is the erbium-induced redistribution of transition energies and oscillator strengths—a dense manifold of mid-energy excitations and a visible-region absorption peak absent in the pristine host. From these spectra the paper selects three scalar descriptors—absorption coefficient alpha, extinction coefficient kappa, and transition energy E—which are then fed to an RBF-kernel support vector machine, a quantum feature-map kernel, and","core_discovery":"The core claim is that dopant-induced changes in the absorption spectrum are a sufficient and physically meaningful signature for classification. Replacing one calcium atom with erbium in a Ca8F16 cluster systematically red-shifts the dominant optical transitions and redistributes oscillator strength, producing enhanced visible absorption near 3 eV and shifting the strongest UV transition from about 7.1 to 6.1 eV. Once the discrete transitions are broadened into continuous spectra, the absorption coefficient, extinction coefficient, and transition energy separate the two systems almost perfectly in a classical kernel model. The same three descriptors, embedded through a quantum feature map,","pith_inferences":["Because the 1,589 energy points per system are densely sampled along two smooth absorption curves, random point-wise train/test splits likely let classifiers interpolate the two known spectra; a harder and more meaningful benchmark would use independent material instances, such as different dopant concentrations, host sizes, or configurations.","The same three descriptors could be tested for transfer to other rare-earth dopants or fluoride hosts; strong transfer would show the fingerprints are dopant-generic, while collapse would reveal that they are specific to the CaF2:Er pair studied here.","The hardware-versus-simulation gap (0.733 vs 0.851) suggests a concrete follow-up test: increasing shot counts or applying error mitigation should move hardware accuracy toward the noisy-simulator value, a prediction that follows from the paper's own noise explanation but is not tested there.","If the aim is to benchmark near-term quantum learning, labeling individual energy bins on two known spectra is easier than classifying unseen material instances; future work should separate these two tasks explicitly."],"forward_implications":["The three selected descriptors carry almost all discriminative information: a classical RBF-kernel SVM achieves 0.983 test accuracy and 0.999 ROC-AUC.","Fixed quantum kernels underperform the classical baseline: 0.851 accuracy on an ideal simulator, 0.817 under simulated depolarizing noise, and 0.733 on real quantum hardware.","A trainable hybrid quantum neural network reaches 0.93 test accuracy and 0.96 AUC, improving on fixed quantum kernels but still below the classical SVM.","The distinguishing spectral features persist across Gaussian broadening widths from 0.1 to 0.2 eV, indicating the fingerprint is not an artifact of the broadening choice.","The end-to-end pipeline—from first-principles spectra to feature selection to quantum kernel execution on hardware—runs successfully and stays above random guessing, even under finite-shot and decoherence constraints."],"fun_headline_variants":["Classical SVM hits 98% on Er-doped CaF2 spectra; quantum lags","Hybrid quantum net scores 93% on dopant-identification task","Optical fingerprints from DFT split CaF2 and Er:CaF2 at 98%","Quantum models trail classical on CaF2 erbium classification","CaF2 erbium fingerprinting: classical 98%, hybrid quantum 93%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the 1,589 energy-resolved points per system are independent, exchangeable samples, even though they are densely spaced points along two smooth absorption curves, so the reported accuracies measure generalization rather than interpolation of the two known spectra.","fun_headline_variants_meta":{"raw":{"variants":["Classical SVM hits 98% on Er-doped CaF2 spectra; quantum lags","Hybrid quantum net scores 93% on dopant-identification task","Optical fingerprints from DFT split CaF2 and Er:CaF2 at 98%","Quantum models trail classical on CaF2 erbium classification","CaF2 erbium fingerprinting: classical 98%, hybrid quantum 93%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2482,"prompt_tokens":894,"completion_tokens":1588,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":1481}},"tokens_in":638,"tokens_out":1588,"duration_ms":16640,"temperature":1.0,"reasoning_tokens":1481,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:58:44.919631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same classifiers on a dataset where the energy grid is coarsely sampled (so neighboring training and test points are not adjacent) or split by contiguous energy blocks; if accuracy collapses or the classical near-perfect separation vanishes, the benchmark is measuring interpolation of two smooth spectra rather than learned discrimination.","supporting_citations":[],"review_version":1}