{"id":"4e3223fa-abb6-4f79-a42d-628c5561d33b","arxiv_id":"2411.10303","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A competition report demonstrating that DMRG, a quantum-inspired solver, scales to 200-ply laminate stacking sequence retrieval and is competitive with classical methods, with small-scale F-VQE results on up to 20 qubits.","lead":"This report describes a quantum-assisted method for stacking sequence retrieval in laminated composite design, a combinatorial step in aerospace optimization. It shows that a tensor-network solver, DMRG, can handle up to 200 plies and beat or match classical reference algorithms on many test cases, suggesting quantum algorithms may one day do the same.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported DMRG advantage is only demonstrated on targets drawn from valid stacking sequences (Sec. 2.3.2); real SSR targets from the first-level continuous optimization are generically off-manifold, so the central 'outperforming' claim is not yet established.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing point: the benchmark targets are all achievable by valid stacking sequences, whereas real retrieval targets from the upper-level continuous optimization are not. The paper's own description of the bi-level workflow (Sec. 2.1) makes this mismatch concrete. I agree with the reader's CONDITIONAL assessment and do not see a reason to change it. The DMRG scalability demonstration (Figs. 8 and 9), the availability of code, and the F-VQE small-instance results are real evidence, so the concern is not that the methods are invalid; it is that the headline comparison has not been tested on the actual distribution of SSR problems. A targeted off-manifold benchmark would settle the question. Because the existing verdict already conditions acceptance on this kind of external validity, I would leave the verdict unchanged rather than escalate to rejection or de-escalate to acceptance.","tokens_in":18433,"tokens_out":4977,"duration_ms":39920,"concrete_test":"Build a new benchmark whose target lamination parameters come from the first-level continuous optimization of a real bi-level run, or, if unavailable, are sampled uniformly from the feasible lamination-parameter region rather than from valid stacking sequences. Run DMRG, LAYLA, and Opti-BLESS with the same active constraints (including Opti-BLESS with constraints enabled, or restrict the comparison to DMRG vs LAYLA if Opti-BLESS becomes impractical) on the same 40 targets per ply count for N=15, 20, 50, 100, 200, and report median Euclidean distance to target and runtime. If DMRG's advantage persists on off-manifold targets, the central claim is supported; if it degrades or reverses, the reported outperformance is an artifact of on-manifold target generation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that with the full set of constraints DMRG outperforms LAYLA and Opti-BLESS on most cases in accuracy and efficiency—is supported only by benchmarks whose target lamination parameters are generated from valid stacking sequences (Sec. 2.3.2: 'we generated a diverse set of target lamination parameters corresponding to valid stacking sequences'). This construction places every target on the discrete, constraint-satisfying manifold and guarantees the optimal loss is 0. In the actual bi-level workflow, the SSR target comes from the first-level continuous lamination-parameter optimization (Sec. 2.1), which generically produces targets that no valid stacking sequence achieves exactly. For such off-manifold targets, accuracy is determined by the trade-off between matching the lamination parameters and satisfying manufacturing constraints; DMRG's penalty weights (Table 2) were tuned on feasible targets where this trade-off is invisible. The comparison is further weakened by Opti-BLESS being run without constraint enforcement (Sec. 2.3.2/5.2.2), so its reported accuracy is not on equal terms. None of this shows the DMRG result is wrong, but it means the central 'outperforming' assertion is an extrapolation from a favorable instance distribution rather than a demonstrated property on representative SSR problems.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports the QAIMS lab's finalist submission to the Airbus/BMW Quantum Computing Challenge. It extends the authors' earlier work on quantum/tensor-network stacking sequence retrieval by (i) enforcing the full set of manufacturing constraints in DMRG, (ii) implementing F-VQE for lamination-parameter matching and buckling-factor maximization, (iii) adding a nearest-neighbor clustering/dispersion term, and (iv) benchmarking DMRG against LAYLA and Opti-BLESS on symmetric laminates with up to 200 plies. The central claim is that, with the full set of constraints, DMRG is competitive with and outperforms LAYLA and Opti-BLESS on most test cases in both accuracy and efficiency, which the authors take as evidence that quantum-inspired and ultimately quantum algorithms could be a viable route for SSR in composite laminate design.","tokens_in":18708,"tokens_out":9162,"duration_ms":86131,"significance":"If the comparative claim survives scrutiny, this would be one of the largest-scale demonstrations of a quantum-inspired method on a practically motivated combinatorial design problem, and the open-source implementation would provide a useful baseline. The DMRG implementation with all manufacturing constraints as MPO penalties, the use of known-optimum benchmark targets, and the explicit head-to-head comparison with two classical SSR codes are concrete strengths. The F-VQE results show that probability-shaping filters maintain superposition and amplify good states, addressing a limitation of the earlier HWE approach. However, the benchmarking limitations identified below mean that the central 'outperforming' conclusion is, at present, an extrapolation from a favorable instance distribution rather than a demonstrated property on representative SSR problems.","major_comments":[{"comment":"The benchmark constructs all target lamination parameters from valid stacking sequences (Section 2.3.2: 'we generated a diverse set of target lamination parameters corresponding to valid stacking sequences'), so every instance has an exactly feasible optimum with loss 0. In the bi-level workflow of Section 2.1, the SSR target comes from the first-level continuous lamination-parameter optimization, which generically produces targets that no valid stacking sequence achieves exactly; for such off-manifold targets, accuracy is determined by the trade-off between matching the lamination parameters and satisfying the manufacturing constraints. The comparison in Figure 8 therefore does not exercise the regime where the claimed advantage matters most. To support the conclusion that DMRG outperforms LAYLA and Opti-BLESS on representative SSR problems, the authors should add benchmarks with off-manifold targets (e.g., perturbed targets or targets from a first-level optimizer) and report both the lamination-parameter distance and the constraint-violation rate.","section":"§2.3.2, Fig. 8"},{"comment":"Opti-BLESS is evaluated without constraint enforcement (Section 2.3.2 and Section 5.2.2), while DMRG and LAYLA produce valid stacks; unconstrained solutions may violate manufacturing rules, so their reported distances are not on equal terms with the DMRG/LAYLA results in Figure 8. The comparison is further weakened because Opti-BLESS was run on only 10 of the 40 targets per configuration (Section 5.2.2), making the distributions noisier. The paper also acknowledges that LAYLA's hyperparameters were not tuned (Section 2.5). The central 'outperforming in most cases' claim needs a like-for-like comparison, with constraints enabled for all methods (or a common feasibility filter) and matched trial counts.","section":"§2.3.2/§5.2.2, Fig. 8"},{"comment":"The buckling-factor experiments do not use a fixed hyperparameter setting: Section 2.4 states that the penalty was 'incrementally increased ... until optimization produced a valid result,' and Section 5.1 reports h=0.1 chosen 'after testing.' Per-trial, outcome-dependent penalty adjustment makes it difficult to separate algorithmic performance from oracle-like tuning. For reproducibility, the constrained buckling results should be obtained under a fixed penalty schedule or a predefined selection rule applied identically to all trials.","section":"§2.4/§5.1/§5.2.1"}],"minor_comments":[{"comment":"The sampling procedure for the 'diverse set of target lamination parameters' is not specified; please state how many random valid sequences were used, how the ply counts were selected, and the random seed, so the benchmark is reproducible.","section":"§2.3.2"},{"comment":"The product in the basis-state encoding should run over |s_n⟩ for n=1,...,N; as typeset it reads |sN⟩ for every factor, which obscures the encoding.","section":"Eq. (7)"},{"comment":"There is a stray comma after (zn−zn−1), and the vD formula uses a summation index k with terms depending on n; please align the indexing and notation.","section":"Eq. (15)"},{"comment":"The caption and Section 5.2.2 do not report how many independent DMRG runs are summarized in each box, nor how the vertical mean-runtime line combines the five bond dimensions and four sweep counts; please clarify.","section":"Fig. 8"},{"comment":"The statement that F-VQE 'consistently identified the optimal state across all trials' should be qualified as 'the optimal basis state was sampled at least once among the 1000 shots,' because Figures 4 and 5 show that the target-state probability remains well below 1 throughout optimization.","section":"§2.4"},{"comment":"The statement that no significant dependence on the parameter-optimization scheme was found is not accompanied by data; either show the comparison or omit the claim.","section":"§5.2.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competition report, and its central comparison is the main weakness. The issues are fixable within the manuscript's scope by extending the benchmark and equalizing the classical baselines; I do not see a need to reject. The paper's self-citation [1] is appropriate given the prior publication. The fit with quant-ph is reasonable, though the contribution is more applied optimization than quantum algorithmics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine extension of the authors' earlier SSR work, and the DMRG comparison is the most useful part. But the headline claim—that DMRG with full constraints outperforms LAYLA and Opti-BLESS on most cases—is not as solid as it looks. The test targets are all generated from valid stacking sequences (Sec. 2.3.2), so the optimal distance is always zero. Real SSR targets from the first-level continuous optimization are generically off that manifold. On those, the trade-off between matching lamination parameters and satisfying constraints matters, and this benchmark never exercises that trade-off. Also, Opti-BLESS was run without constraint enforcement (Sec. 5.2.2), which is not a fair comparison; the authors say so, but it means the 'accuracy' comparison is apples-to-oranges. LAYLA hyperparameters were not tuned, which they also acknowledge.\n\nWhat is genuinely new: full manufacturing constraints encoded in the DMRG MPO, the F-VQE implementation (limited to N=8,10 but works), the 15-degree ply-angle demonstration, the buckling objective, and the clustering/dispersion bias. The flexibility claims are credible. The code is on GitHub, and the DMRG scaling to 200 plies with stable error around 0.2 is a real data point. I believe the DMRG results are probably reproducible and the comparison is honestly reported, aside from the Opti-BLESS caveat.\n\nSoft spots in proportion: the on-manifold target issue is the biggest one. It doesn't disprove the method, but it means the 'outperforming' statement is an extrapolation from a favorable instance distribution. The buckling experiments involved per-trial penalty adjustment until success, which is a free parameter issue. No confidence intervals or error bars on the box plots; with 40 targets and sometimes only 10 for Opti-BLESS, the medians could shift. The F-VQE results are small and use 1000 shots, so they are a proof-of-concept, not evidence of quantum advantage. The 'conservative baseline for quantum advantage' claim is speculative—DMRG's entanglement restrictions don't guarantee quantum hardware would do better.\n\nOverall: this is a well-written competition report, not a definitive study. It deserves a serious referee, but the referee should ask for tests on off-manifold targets and a constrained Opti-BLESS run before the central claim is accepted. I'd cite it for the DMRG benchmark and the constraint MPO if I worked on SSR; I wouldn't cite the quantum-advantage language.","headline":"Solid competition report with a credible DMRG benchmark, but the 'outperforming classical methods' claim rests on on-manifold test targets and an unfair Opti-BLESS baseline, so treat the headline as plausible rather than proven.","tokens_in":19270,"tokens_out":2619,"would_cite":true,"duration_ms":23861,"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":"The paper claims that stacking sequence retrieval in composite design can be solved by quantum-assisted methods, and that a tensor-network solver matches or beats the classical state of the art on most tested cases up to 200 plies.","keywords":["stacking sequence retrieval","lamination parameters","composite laminates","DMRG","F-VQE","quantum optimization","manufacturing constraints","tensor networks"],"falsifier":"Run DMRG, LAYLA, and Opti-BLESS on random target points inside the feasible lamination-parameter space rather than targets derived from valid sequences, holding ply count and constraints fixed. If DMRG's error and runtime advantage over LAYLA and Opti-BLESS shrinks or reverses on such targets, the paper's competitiveness claim is specific to its generated test distribution.","tokens_in":18227,"feed_emoji":"⚛️","tokens_out":8126,"duration_ms":76583,"temperature":0.7,"pith_summary":"Stacking sequence retrieval — choosing the discrete ply-angle sequence so that a composite laminate's stiffness matches target lamination parameters — is the combinatorial bottleneck in the standard bi-level optimization of composite structures. This paper claims that the problem can be cast as a ground-state search on a diagonal quantum Hamiltonian and solved with a quantum-inspired tensor-network solver (DMRG) and a variational quantum algorithm (F-VQE). In a comparison with two established classical solvers, LAYLA and Opti-BLESS, DMRG outperformed them on most test cases in both accuracy and runtime while enforcing the full set of manufacturing constraints, up to 200 plies. F-VQE found the optimal stacking sequence in every trial for 8- and 10-ply laminates and, unlike the earlier hardware-efficient circuit, kept a superposition of candidate states during optimization. The same framework handles finer 15-degree ply-angle increments, buckling-factor maximization, and a tunable penalty for clustering or dispersing same-angle plies.","feed_headline":"Tensor-network solver beats classical layup search on most cases","feed_subtitle":"Quantum-inspired solver matches target stiffness and beats two established layup-design methods.","key_machinery":"The load-bearing construction is a diagonal Hamiltonian whose basis states encode ply-angle sequences: each ply is a $d$-dimensional quantum register (two qubits for the four conventional angles), and the energy of a state is the squared distance from the target lamination parameters plus penalty terms for disorientation, contiguity, the 10% rule, and the balanced condition. DMRG minimizes this Hamiltonian in matrix product form, with bond dimension and number of sweeps as the primary tuning parameters. F-VQE applies a filter function to the measured probability distribution so that low-energy states are amplified while the overall state remains a superposition. A nearest-neighbor bias term $\\alpha \\sum_{n=1}^{N-1} \\delta_{s_n,s_{n+1}}$ tunes clustering versus dispersion of same-angle plies, and an excitation-preserving partial-swap circuit conserves ply-angle counts when the objective is, for example, maximizing the buckling factor under biaxial loading.","core_discovery":"The central claim is that stacking sequence retrieval can be represented as the minimization of a Hamiltonian over quantum basis states, with manufacturing constraints added as penalty terms, and that this representation is not just a theoretical construction but a working solver. DMRG, run as a classical tensor-network ground-state solver, produced valid stacking sequences whose lamination parameters stayed within about 0.2 of the targets for ply counts from 15 to 200, and it required less computation time than LAYLA on nearly all configurations while matching or beating both classical methods on accuracy. F-VQE, which reshapes the probability distribution of the variational state to amplify low-energy configurations, is claimed to fix the early state-collapse observed in the authors' earlier variational approach and to converge to the optimal stacking sequence across all tested 8- and 10-ply instances. Because DMRG cannot represent the entanglement a true quantum computer could generate, the authors present it as a conservative baseline and argue that actual quantum hardware is likely to do at least as well.","pith_inferences":["A next natural test, not reported in the paper, would be to run the same comparison on targets produced by a real first-level optimizer, since targets generated from valid sequences make the optimal loss exactly zero by construction.","The near-constant error of about 0.2 across ply counts suggests the bottleneck is the solver's representation, not the problem size; if so, even larger laminates are within reach at that accuracy.","The same basis-state encoding applies to any layer-wise design problem with discrete choices and additive layer contributions, so the method could be carried over to wind-turbine blades, battery stacks, or hybrid-material panels."],"forward_implications":["If these results are correct, a tensor-network solver with modest bond dimensions is a practical alternative for the stacking sequence retrieval step in composite design, since it stays accurate to about 0.2 in lamination-parameter distance up to 200 plies.","The F-VQE demonstration implies that variational quantum circuits can avoid premature collapse and amplify good candidate states, addressing a known weakness of earlier hardware-efficient variational approaches.","Handling 15-degree ply-angle increments opens the formulation to design spaces beyond the conventional 0, ±45, and 90 degree laminates without changing the underlying encoding.","The buckling-factor objective and the nearest-neighbor clustering bias show that the same machinery can optimize properties beyond stiffness matching and can encode manufacturing-cost preferences.","Since DMRG is entanglement-limited, the authors conclude that actual quantum algorithms are a potential upgrade path that could exceed this baseline performance."],"supporting_citations":[{"why":"Supplies the basis-state encoding, Hamiltonian formulation, constraint penalties, and DMRG sweep strategy that this work extends to the full constraint set.","marker":"[1]"},{"why":"Introduces F-VQE, the filtering variational algorithm whose probability shaping replaces the simpler hardware-efficient circuit from the prior work.","marker":"[23]"},{"why":"Defines the LAYLA beam-search method and the manufacturing constraint set used as the comparison baseline.","marker":"[24]"},{"why":"Is the LAYLA software library that the paper runs as the beam-search baseline.","marker":"[33]"},{"why":"Is the Opti-BLESS genetic-algorithm library used as the second classical baseline.","marker":"[34]"},{"why":"Provides the repair algorithm that keeps LAYLA's stacking sequences constraint-valid, which matters for the accuracy comparison.","marker":"[35]"},{"why":"Is the original DMRG algorithm, the tensor-network solver that carries the large-scale results up to 200 plies.","marker":"[29]"}],"fun_headline_variants":["Quantum-inspired layup solver beats classical search","Tensor network cracks composite stacking problem","DMRG solver outperforms classical layup methods","Quantum-assisted design outsmarts classical stackers","Layup optimization gets a quantum boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that target lamination parameters generated from valid stacking sequences represent the targets that actually arise from the first level of bi-level optimization; if real targets are less regular or not achievable by any valid sequence, the reported accuracy and runtime comparisons may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-inspired layup solver beats classical search","Tensor network cracks composite stacking problem","DMRG solver outperforms classical layup methods","Quantum-assisted design outsmarts classical stackers","Layup optimization gets a quantum boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1365,"prompt_tokens":1005,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":621,"tokens_out":360,"duration_ms":4356,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:45:30.272934+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DMRG, LAYLA, and Opti-BLESS on random target points inside the feasible lamination-parameter space rather than targets derived from valid sequences, holding ply count and constraints fixed. If DMRG's error and runtime advantage over LAYLA and Opti-BLESS shrinks or reverses on such targets, the paper's competitiveness claim is specific to its generated test distribution.","supporting_citations":[{"cited_title":"A method using beam search to design the lay-ups of composite laminates with many plies","cited_arxiv_id":null,"evidence_quote":"Defines the LAYLA beam-search method and the manufacturing constraint set used as the comparison baseline."},{"cited_title":"LAYLA: a method using beam search to design the lay-ups of composite laminates with many plies","cited_arxiv_id":null,"evidence_quote":"Is the LAYLA software library that the paper runs as the beam-search baseline."},{"cited_title":"Opti-BLESS","cited_arxiv_id":null,"evidence_quote":"Is the Opti-BLESS genetic-algorithm library used as the second classical baseline."}],"review_version":1}