{"id":"118a767d-0909-4c71-86cb-72208d32c4c2","arxiv_id":"2507.10237","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Quantum annealing based QBoost and QSVM models classify six high-entropy alloy phases with accuracy comparable or superior to classical SVM in several cases, but with a flawed runtime benchmark and unresolved label inconsistencies.","lead":"This paper applies quantum annealing to machine learning models that predict which crystal phase a high-entropy alloy will form. It reports that the quantum versions match or beat classical models on some phases and run faster in their setup, but the speedup and accuracy claims are not fully supported by the data as presented.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported accuracies rest on an undefined phase-presence label: Table 4 marks many multiphase alloys (e.g., B2+A1) as FALSE for FCC/BCC despite those phases being present, so the headline 'match or exceed classical' lacks a reproducible target.","rationale":"I focused on the label ambiguity rather than the runtime or QSVM-constraint issues because it is the least caveated and most fundamental. The runtime advantage is explicitly scoped as hardware/problem-dependent ('in our setup'), and the QSVM discussion openly describes relaxing the equality constraint as an intentional regularizer; both are debatable but at least stated. In contrast, the paper asserts that datasets capture presence/absence of each phase, yet Table 4—the paper's own per-alloy test record—contains systematic mismatches under that semantics. If the ground-truth labels mean something else, the abstract's 'match or exceed classical models in accuracy' cannot be evaluated. The reader's verdict of REJECT is therefore unchanged by this pass; the label ambiguity alone is sufficient to block acceptance as stated, though a clarified relabeling and recomputation could make the underlying pipeline salvageable.","tokens_in":17163,"tokens_out":7934,"duration_ms":94563,"concrete_test":"Recompute per-phase test accuracy and F1 from Table 4 under the presence semantics (A1→FCC, A2→BCC, any Laves→Laves+, any Sigma→Sigma+, L2₁→Heusler+, any B2→B2+), then under a majority/single-phase semantics, and compare both to the testing rows of Table 2 and Table 3. If neither semantic reproduces the published numbers within rounding, the label ambiguity is confirmed. As a supplementary check, obtain the exact per-alloy label files from the authors and recompute the confusion matrices for the FCC and BCC models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's Data Construction states that each dataset 'was constructed to capture both presence and absence of the phase of interest.' Under that semantics, an alloy whose XRD shows A1 anywhere should be labeled True for the FCC model, and one showing A2 should be True for BCC; B2+Laves should be True for Laves; L2₁ should be True for Heusler. Table 4 does not follow this. Examples: Ag20Al20Cr20Mn20Ni20 (B2+A1) is FCC=False; Al15Cr15Mo15Ni46W9 (B2+A1+A2) is FCC=False and BCC=False; Al33Co17Nb33Ni17 (B2+Laves) is Laves=False; Al25Cr10Fe20Mn10Ni20Ti15 (L2₁) is Heusler=False. If 'presence' is the target, these are false negatives and the reported test accuracies in Tables 2 and 3 are not reproducible from the table. If the target is instead 'dominant single phase' or 'A1 as a majority phase,' that must be stated and the feature-selection and accuracy claims re-interpreted. Because every accuracy, F1, and interpretability conclusion is computed against these labels, this ambiguity is load-bearing: resolving it in favor of presence would invalidate the reported metrics, while resolving it toward dominance would require re-deriving all headline comparisons.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantum-annealing-based machine learning pipeline for classifying six crystallographic phases (FCC, BCC, Al-X-Y B2, Heusler, Sigma, Laves) in high-entropy alloys. The pipeline combines QBoost for interpretable feature selection and classification, and a QSVM for nonlinear classification, both formulated as QUBO problems and solved on a D-Wave annealer (with the hybrid solver for QSVM). The authors benchmark against simulated annealing and classical SVM, using cross-validation plus an independent test set of 86 experimentally synthesized alloys from a prior study, and report competitive accuracy, interpretable feature selection, and several orders of magnitude runtime reduction.","tokens_in":17538,"tokens_out":5401,"duration_ms":60983,"significance":"If substantiated, the work would be a practical demonstration of quantum annealing in an interpretable materials-informatics pipeline, with useful feature-selection insights and a real experimental test set. The core QUBO derivations for QBoost follow established work (Neven et al., Willsch et al.) and are clearly presented. The use of an independent experimental test set is a strength. However, the central empirical claims are not currently supported: the phase-label semantics underlying every accuracy number are ambiguous and inconsistent with Table 4; the QSVM is not a standard SVM because the equality constraint is dropped; and the runtime comparison uses an implausibly slow simulated annealing baseline. These issues are load-bearing for the headline conclusions of accuracy, generalization, and quantum advantage.","major_comments":[{"comment":"The label definition is ambiguous and internally inconsistent. The Methods section states that each dataset 'was constructed to capture both presence and absence of the phase of interest.' Under that semantics, an alloy whose XRD shows A1 (FCC) anywhere in a multiphase assemblage should have a positive FCC label, and one showing A2 should have a positive BCC label. Table 4 contradicts this: for example, Ag20Al20Cr20Mn20Ni20 (experimental B2+A1) is listed with FCC prediction FALSE; Al15Cr15Mo15Ni46W9 (B2+A1+A2) has both FCC and BCC FALSE; Al33Co17Nb33Ni17 (B2+Laves) has Laves FALSE; Al25Cr10Fe20Mn10Ni20Ti15 (L2₁, a Heusler phase) has Heusler FALSE; and Al10Co20Cr10Cu20Mn20Ni20 (A1+A2) has FCC TRUE but BCC FALSE. If 'presence' is the target, these are false negatives and the test accuracies in Tables 2 and 3 are not reproducible from the table. If the target is instead something like 'dominant phase' or 'single-phase presence,' that must be stated explicitly, and all reported accuracies, F1 scores, and the interpretability conclusions would need to be recomputed under that definition. Because every quantitative claim in the paper depends on this label choice, this ambiguity is load-bearing.","section":"Data Construction; Table 4"},{"comment":"The QSVM is not a support vector machine as defined, because the equality constraint sum_n alpha_n y_n = 0 is not enforced in the best-performing models. Equation (3)-(4) define the dual SVM problem with constraint sum alpha_n y_n = 0; the QUBO in Methods includes a penalty term xi*(sum alpha_n y_n)^2, but Table 3 shows that all reported best QSVM models use xi = 0. The paper itself acknowledges that 'the best-performing QSVM models emerged when the constraint was relaxed rather than strictly enforced.' Without this constraint, the bias term b in the decision function is not well defined by the standard SVM formula given in Methods, since the denominator sum alpha_n (C - alpha_n) can vanish or the KKT conditions do not hold. The classifier is therefore a different model (a regularized kernel classifier with stochastic coefficients), and the comparison with a classical SVM that enforces the constraint is not a controlled comparison. To support the claim that quantum annealing provides a benefit, the authors should either enforce the constraint (nonzero xi) and report those results, or compare against a classical model that also relaxes the constraint in the same way.","section":"Quantum Support Vector Machine; Methods equation for bias"},{"comment":"The reported runtime advantage of roughly 10,000x is based on an unfair baseline. The paper states that QA solved each QUBO in approximately 0.01-0.02 seconds, while SA took about 300-400 seconds per instance on a CPU cluster. For a QUBO with only 31 binary variables and 1000 samples (the stated QA parameters), a well-implemented simulated annealer on a modern CPU should finish in milliseconds or at most a few seconds; 300-400 seconds suggests a pathological implementation or a mismatched workload (e.g., running SA on a dense, fully connected QUBO with default parameters that do not exploit sparsity). The Methods only says SA used 'D-Wave's reference implementation from Ocean SDK, using dimod package, matching the QA sample count and using default parameters,' which is not sufficient to establish a fair time-to-solution comparison. The paper's own caveat that speedups are hardware-dependent is appropriate, but the figure and the phrase 'orders-of-magnitude faster convergence' are not supported by the evidence as presented. The authors should benchmark SA with a standard, optimized implementation and report time-to-solution as a function of problem size, not just wall-clock time for a single instance.","section":"Results; Methods - Simulated Annealing"}],"minor_comments":[{"comment":"The claim that the framework 'consistently delivers competitive, and often superior, performance' is overstated relative to Table 3: classical SVM achieves higher test accuracy than QSVM for FCC (0.93 vs 0.90) and BCC (0.95 vs 0.92), equal for Laves (0.90), and higher for none of the remaining phases in a consistent pattern; the advantage is phase-dependent.","section":"Discussion"},{"comment":"There are copy-paste errors in the dataset descriptions: the FCC dataset is said to comprise '132 HEAs with the FCC phase (Laves)' and the BCC dataset '178 HEAs with the FCC phase (Laves)'; these should refer to the FCC and BCC phases respectively.","section":"Data Construction"},{"comment":"The entry 'Co18Cu18Fe18Mn18Ni18V10' lists experimental result 'A1+A1', which appears to be a typo (likely A1+A2 or another mixture); this should be corrected.","section":"Table 4"},{"comment":"The reference numbering is inconsistent: reference [32] is cited as both 'Pudenz & Lidar (2013)' and later 'Brown & Zhuang (2023)' in the same paragraph, and some references (e.g., [34]) have mismatched arXiv numbers; the reference list should be carefully reconciled.","section":"Introduction"},{"comment":"The caption of Figure 5 does not specify which dataset or hyperparameters were used beyond 'RB2 dataset'; for reproducibility, the exact parameter settings (B, K, xi, gamma, number of reads) should be given.","section":"Figure 5"},{"comment":"The abstract states 'rigorously curated test set' but the test set is drawn from the authors' prior study [20]; this is acceptable, but the text should clarify that these are the same compositions and that no new experimental data are reported here.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The label ambiguity in Table 4 is, in my view, disqualifying: it affects every reported accuracy, F1, and feature-selection claim, and it is not a minor presentation issue. The QSVM equality-constraint issue and the unfair SA baseline independently undermine the two other headline claims (quantum advantage in accuracy and in runtime). Even though these issues are in principle fixable, the necessary re-benchmarking and re-analysis would change the paper's central conclusions, so I cannot recommend major revision as a path to accept. The paper does contain useful QUBO formulations and a clear presentation of the QBoost methodology, but the empirical validation as written does not support the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first application of quantum-annealing-based QBoost and QSVM to HEA phase classification, and the relaxed-constraint observation is worth a second look. But the reported test accuracies don't survive contact with the paper's own Table 4.\n\nWhat's new: applying established QUBO-based QBoost and QSVM to a materials problem, with an independent experimental test set of 86 alloys. The feature-selection heatmap and the empirical finding that dropping the SVM equality constraint acts as a regularizer are genuine contributions. The QUBO derivations are standard, but they're assembled cleanly.\n\nSoft spots, in order of severity:\n\n1. The label semantics. Data Construction says each dataset captures \"presence and absence\" of the phase. Table 4 lists experimental results like \"B2+A1\" and yet marks FCC as False for those alloys. If the task is presence, those are false negatives, and the reported FCC/BCC accuracies in Tables 2 and 3 don't match the table. If the task is dominant-phase or something else, it needs to be stated and all metrics recomputed. This is load-bearing, because every claim in the abstract depends on it.\n\n2. The runtime comparison. SA taking 300–400 seconds for a 31-variable QUBO is implausible for a decent implementation, and the QA times are hardware-specific. The paper concedes hardware-dependence, but the \"10,000x\" figure will mislead if the SA baseline isn't tightened.\n\n3. QSVM's bias term. The paper correctly notes that the equality constraint is needed for a well-defined bias, but all best models use xi=0 and no bias derivation is given for the relaxed case. It's an interesting finding, but the decision function is underspecified.\n\nWho gets value: readers working on quantum annealing for materials informatics will find the pipeline template useful, but the results need re-evaluation after the label issue is resolved.\n\nMy call: this deserves a serious referee because the application is new and the flaws are repairable. But I wouldn't cite it until the label question is answered, and I'm not convinced the current accuracy claims are correct. Send it back for major revision.","headline":"A genuinely new application of QA-based QBoost/QSVM to HEA phase classification, but the label ambiguity in Table 4 undermines the headline accuracy claims.","tokens_in":18024,"tokens_out":3656,"would_cite":false,"duration_ms":38789,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum annealing provides competitive, interpretable phase classification for high-entropy alloys, with faster training in the authors' setup.","keywords":["quantum annealing","high-entropy alloys","phase classification","QUBO","QBoost","quantum support vector machine","interpretable machine learning","materials informatics"],"falsifier":"A reader can settle the generalization claim by scoring Table 4 under the natural rule 'phase present anywhere in the XRD assemblage counts as positive' and recomputing the per-phase test accuracies; for FCC the listed A1-containing multiphase alloys predicted FALSE would push the accuracy below the reported 0.8833. If the authors instead meant single-phase dominance, the same table needs a label definition before any reported accuracy is reproducible.","tokens_in":16980,"feed_emoji":"⚛️","tokens_out":8979,"duration_ms":96541,"temperature":0.7,"pith_summary":"The paper tries to establish that quantum annealing can run two standard machine-learning models—boosting and support vector machines—by casting their training as quadratic unconstrained binary optimization (QUBO), and that the resulting classifiers are practically useful for high-entropy alloys. Applied to six target phases (FCC, BCC, Sigma, Laves, Heusler, and Al–X–Y B2), the QBoost and QSVM models match or exceed classical baselines while selecting physically interpretable descriptors such as valence electron concentration and mixing enthalpy. If the claim is right, alloy designers gain a faster, explainable screening tool in a regime where data are scarce and phase labels come from X-ray diffraction. The paper also reports that relaxing the SVM equality constraint improves test-set generalization, which it interprets as stochastic regularization from the annealer's near-optimal solutions.","feed_headline":"Quantum annealers match classical ML on alloy phase prediction","feed_subtitle":"QBoost and QSVM select physically meaningful descriptors and train orders of magnitude faster than simulated annealing here.","key_machinery":"The load-bearing object is the QUBO reformulation of both learners. For QBoost, the squared boosting loss over binary weights on single-feature decision stumps (thresholds on one physical descriptor) expands into a quadratic form whose coupling matrix is $C'_{ij} = \\sum_s h_i(x_s) h_j(x_s)$; switching to spin variables turns it into an Ising Hamiltonian, so each feature's inclusion is one spin. For QSVM, each dual coefficient $\\alpha_n$ is expanded in base $B$ with $K$ bits, and the SVM dual loss, plus a penalty $\\xi (\\sum_n \\alpha_n y_n)^2$ for the equality constraint, becomes a QUBO matrix $Q$ with entries $B^k B^j y_n y_m [K(x_n, x_m) + \\xi]$ minus the diagonal term; the annealer minimizes this matrix. The same binary-optimization architecture carries both models, and the interpretability comes from the one-descriptor-per-stump structure and from keeping features in their original physical units.","core_discovery":"The central claim is that high-entropy alloy phase classification can be reformulated as QUBO and solved on current quantum annealing hardware without giving up accuracy. QBoost trains an ensemble of single-feature decision stumps with binary weights, so the optimized weight vector is simultaneously a classifier and a feature selector; QSVM encodes support-vector dual coefficients in binary and uses an RBF kernel, so the SVM dual loss becomes a QUBO matrix that the annealer minimizes. On five-fold cross-validation and on an independent test set of 86 experimentally synthesized and XRD-characterized alloys, the paper reports that quantum-annealed models match or beat simulated-annealing-selected models and classical SVMs on most phases, with QSVM particularly improving test accuracy for the harder intermetallic phases (RB2 test accuracy 0.93 versus 0.86 for the previous classical model). The authors further claim that the best QSVM models are obtained when the constraint $\\sum_n \\alpha_n y_n = 0$ is relaxed rather than strictly enforced, turning the annealer's stochastic near-optimal solutions into an implicit regularizer.","pith_inferences":["The relaxed-constraint result points to a design rule for QUBO-based machine learning generally: deliberately leaving a KKT-type equality unsatisfied may act like dropout or early stopping. Testing the same penalty sweep on other datasets would show whether the effect is specific to this kernel and feature set.","The runtime advantage is a hardware-and-implementation comparison, not a proven quantum speedup; a cleaner benchmark would measure end-to-end wall-clock time including embedding, against highly optimized classical QUBO solvers and classical SVMs.","The Table 4 label ambiguity is the main reproducibility risk: if 'FCC FALSE' means the phase is not the dominant phase, rather than absent, then the reported 0.8833 test accuracy is not what a reader would compute by checking for A1 in each XRD assemblage. Publishing an explicit label definition or per-prediction error table would settle it.","The selected-feature patterns could be compared against classical sparse classifiers on the same 31 descriptors to isolate what the quantum optimizer adds beyond sparsity, since QBoost's interpretability largely comes from the decision-stump structure rather than from quantum annealing itself."],"forward_implications":["QBoost can act as a transparent feature selector, so a phase prediction comes with a short list of physical descriptors whose role in phase stability is already known.","The adaptive pipeline means simple alloys can be handled by the compact QBoost model, while hard intermetallic phases are passed to QSVM with a physics-grounded feature subset.","Because the reported QUBO solve times are orders of magnitude below the simulated-annealing baseline in this setup, retraining-heavy workflows such as active learning become more practical if the hardware overhead is included in the comparison.","Relaxing the SVM equality constraint consistently improved test generalization in the paper's experiments, so the annealer's approximate solutions should be treated as a feature, not a bug.","The method improves over the earlier feature-engineered model for Laves, Sigma, and RB2 but not for Heusler, marking the limit of the current descriptor set."],"supporting_citations":[{"why":"Supplies the 86 experimentally synthesized HEA test alloys with XRD phase labels and the previous logistic-regression model that the quantum pipeline is benchmarked against.","marker":"[20]"},{"why":"Introduces QBoost and the QUBO formulation of boosting on binary classifier weights, which the paper follows to construct its Ising Hamiltonian.","marker":"[30]"},{"why":"The companion QBoost paper that supplies the adiabatic-quantum-optimization derivation for training a binary classifier, used here for the QUBO construction.","marker":"[31]"},{"why":"Provides the QSVM-on-quantum-annealer methodology, including binary encoding of dual coefficients and the observation that quantum-annealing noise acts as an implicit regularizer.","marker":"[29]"},{"why":"Demonstrates quantum and classical annealing-based boosting reaching accuracy competitive with DNNs and XGBoost, the prior evidence that QA-boosted models are worth applying to HEA data.","marker":"[28]"},{"why":"Recent implementation of QBoost on quantum annealing hardware that grounds the paper's practical QBoost workflow.","marker":"[33]"},{"why":"Probabilistic QSVM training on coherent Ising machines, cited as parallel evidence that stochastic quantum optimization regularizes SVM training.","marker":"[34]"},{"why":"Supplies the phase-diagram heuristics used to build the 31 physics-informed descriptors fed to QBoost.","marker":"[18]"},{"why":"Establishes the valence-electron-concentration rule for FCC/BCC stability that the paper invokes to explain why cubic phases are easiest to classify.","marker":"[40]"},{"why":"The reference the authors invoke for why their runtime speedup is hardware-dependent and not a proven quantum speedup.","marker":"[54]"}],"fun_headline_variants":["Quantum annealing speeds alloy phase prediction without accuracy loss","QBoost and QSVM match classical models on HEA phases","Quantum annealers tip the scales on alloy phase classification","Faster quantum ML for high-entropy alloy phases","Quantum annealing boosts interpretable alloy phase classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training and test labels mean the same thing for each phase—specifically, whether a phase counts as present when it appears in a multiphase X-ray diffraction assemblage—and Table 4 suggests this may not hold for FCC, so every reported accuracy depends on that labeling convention.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealing speeds alloy phase prediction without accuracy loss","QBoost and QSVM match classical models on HEA phases","Quantum annealers tip the scales on alloy phase classification","Faster quantum ML for high-entropy alloy phases","Quantum annealing boosts interpretable alloy phase classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2950,"prompt_tokens":989,"completion_tokens":1961,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":1895}},"tokens_in":605,"tokens_out":1961,"duration_ms":14567,"temperature":1.0,"reasoning_tokens":1895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:37:37.664662+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader can settle the generalization claim by scoring Table 4 under the natural rule 'phase present anywhere in the XRD assemblage counts as positive' and recomputing the per-phase test accuracies; for FCC the listed A1-containing multiphase alloys predicted FALSE would push the accuracy below the reported 0.8833. If the authors instead meant single-phase dominance, the same table needs a label definition before any reported accuracy is reproducible.","supporting_citations":[],"review_version":1}