{"id":"ddb624b2-ec6f-4a66-9862-f455e5ce05c4","arxiv_id":"2411.14259","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Quantum neural networks can classify shock and turbulent flow solutions encoded as quantum states, with accuracy strongly dependent on Fourier versus real-space basis choice.","lead":"This paper tests whether quantum neural networks can classify solutions of fluid dynamics equations directly from quantum states, avoiding the expensive process of reading out the full solution. The results show high accuracy on two small simulated benchmarks, but the performance depends strongly on the measurement basis and no comparison with classical methods is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's core equivalence assumption—that classically amplitude-encoded PDE solutions reproduce noiseless quantum solver outputs—is asserted, not demonstrated, and is load-bearing for the readout claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the paper's demonstration relies on classical amplitude-encoded PDE solutions standing in for actual quantum solver outputs. I agree that this is the central point on which the readout claim depends. The paper explicitly asserts the equivalence in the Results section, but no evidence is provided that a concrete quantum PDE solver for these nonlinear equations produces states matching the classically computed, noise-augmented amplitude vectors. If the equivalence fails, the learned QNN measurement operators may not work on real quantum data, so the readout problem is not actually addressed. If the equivalence holds, the demonstration would be a valid proof-of-concept. The concern is not an internal inconsistency but an unsupported, load-bearing assumption; it warrants a conditional verdict rather than rejection. The absence of classical baselines and reproducible code is secondary because it affects confidence in the numbers, not the logical structure of the claim. My recommended verdict is therefore unchanged from the reader's CONDITIONAL, pending the proposed test.","tokens_in":13250,"tokens_out":4236,"duration_ms":47312,"concrete_test":"Implement a small noiseless quantum PDE solver for the viscous Burgers equation—e.g., the quantum iterative method from ref. [44] or a Carleman-linearized LCU solver—on 6–8 qubits. Generate the same set of shock/steady test cases as output quantum states, then retrain the DQNN with identical hyperparameters on those solver-output states and evaluate on held-out solver states. If the test accuracy falls below the claimed 100% (or below 90%), the equivalence assumption is falsified; if accuracy matches, the proof-of-concept transfers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that quantum scientific machine learning can address the readout problem by learning measurement operators from quantum data generated by quantum PDE solvers. The demonstration, however, never runs a quantum PDE solver. It amplitude-encodes classically computed Burgers solutions (with added white noise) and Navier-Stokes lattice Boltzmann snapshots, then trains QNNs on those states. The Results section states: 'The results from the analysis presented here are the same as if the quantum data inputs were measured directly from a noiseless quantum PDE solver without any classical processing.' This equivalence is assumed, not proven. For the nonlinear PDEs considered (Burgers and Navier-Stokes), concrete quantum solvers typically rely on linearization techniques such as Carleman embedding or Schrodingerisation, producing states in enlarged Hilbert spaces with truncation errors, not necessarily the simple normalized grid-function state used here. If actual solver output states differ in structure, entanglement, or normalization, the trained measurement operators may not transfer, and the claimed solution to the readout problem would fail in the intended application. Additionally, training is performed classically with exact statevector expectation values, so shot noise, measurement sampling overhead, and the cost of estimating the covariance hypothesis (which requires multiple expectation values) are not assessed. The absence of classical baselines and error bars further weakens the evidence, but the equivalence assumption is the most load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a quantum scientific machine learning (QuaSciML) pipeline to address the readout problem for quantum differential equation solvers. It treats solver outputs as quantum data and trains parametrized quantum circuits—a deep QNN for Burgers shock detection and a double-QCNN for Navier-Stokes laminar versus turbulent classification—to act as problem-specific measurement operators that output binary labels directly from quantum states. Training is performed on amplitude-encoded classical simulations of Burgers and Navier-Stokes solutions, with variants using real-space and Fourier bases and with or without coarse-graining. The paper reports 100% accuracy for shock detection with the Fourier plus coarse-grained model and about 92% and 87% accuracy for turbulence classification in real and Fourier bases, respectively, and concludes that learned measurement operators can bypass tomography-based readout.","tokens_in":13519,"tokens_out":4624,"duration_ms":45600,"significance":"If the central equivalence assumption held, the paper would be a useful proof-of-concept that quantum machine learning can serve as a readout layer for quantum PDE solvers, reducing high-dimensional quantum states to low-dimensional classification labels without full tomography. The framing of quantum data as a resource, the emphasis on basis choice and coarse-graining, and the physics-informed covariance hypothesis are valuable ideas. The paper also draws on a relevant body of literature on QCNNs and quantum data learning. However, the demonstration is a classical simulation study with an asserted rather than verified connection to actual quantum solver outputs, and the statistical support for the headline results is limited. The strengths are conceptual and methodological; the experimental evidence is not yet sufficient for the central claim as stated.","major_comments":[{"comment":"The central claim rests on the assertion, 'The results from the analysis presented here are the same as if the quantum data inputs were measured directly from a noiseless quantum PDE solver without any classical processing.' This equivalence is not demonstrated. For the nonlinear PDEs considered, standard quantum algorithms (e.g., Carleman embedding, Schrodingerisation) prepare states in enlarged Hilbert spaces with truncation errors and normalization effects, which need not coincide with the simple amplitude-encoded grid functions used here. If the actual solver states differ in structure or entanglement, the trained measurement operators may not transfer, and the claimed solution to the readout problem would fail in the intended application. Please either run a concrete quantum PDE solver on small instances and test the trained models on its output states, or explicitly restrict the claim to amplitude-encoded quantum states and justify why this restricted setting is representative.","section":"Results, paragraph beginning 'For simplicity...'"},{"comment":"The 100% shock-detection accuracy is obtained from only 22 solutions (11 per class), and the optimal model is selected among the three variants (real, Fourier, Fourier plus coarse-graining) after examining the full-data accuracy curves. The accuracy plotted in Fig. 2(c) is evaluated on the entire data set after training on a subset, which includes training points, so it is not a held-out generalization measure; the confusion matrix in Fig. 2(d) uses an 11-sample test set. With this sample size, 100% accuracy has very wide confidence intervals, and post-hoc model selection without a separate validation set inflates the reported performance. Please provide confidence intervals, a nested cross-validation or a pre-specified model selection rule, and a larger number of independent solutions.","section":"Fig. 2(c,d) and surrounding text"},{"comment":"No classical baseline is provided for either benchmark. A classical machine learning classifier operating on the same truncated and subsampled features, or on the amplitude vectors, is needed to determine whether the quantum measurement layer is actually necessary or whether the tasks become trivially separable after the coarse-graining and late-time truncation. Without such a baseline, the claim that QuaSciML 'addresses the readout problem' is not established beyond showing that a particular quantum circuit can fit this dataset.","section":"Results (Navier-Stokes) and Fig. 4"},{"comment":"Training and evaluation use exact statevector expectation values, so the measurement overhead required to estimate h(ρ,θ) and cov(O1,O2) on actual hardware is not analyzed. Since the readout problem is specifically about the cost of extracting information from quantum states, an analysis of the number of shots needed to estimate these expectation values to the accuracy required for the reported classification results is essential. Without it, the proposal may simply shift the tomographic cost into repeated measurements of a covariance hypothesis.","section":"Hypothesis definitions and Eq. (4)"}],"minor_comments":[{"comment":"The text states that Burgers solutions are encoded into 12-qubit states, while the Fig. 2 caption says 6-qubit quantum states; please reconcile these values.","section":"Fig. 2 caption and Burgers results text"},{"comment":"The terms 'double-QNN' and 'double-QCNN' are used interchangeably; please choose one consistent name for the two-register architecture.","section":"General notation"},{"comment":"Fig. 4(b,d) report results over four random seeds, while Fig. 4(c,e) report averages over ten seeds; please make the number of seeds consistent or explain the discrepancy.","section":"Fig. 4(b-e)"},{"comment":"Reference [96] appears to lack author information and a full citation; please complete it.","section":"Reference [96]"},{"comment":"The color legend for steady waves versus shock waves is not explicitly defined in the caption; please add a clear legend.","section":"Fig. 2(a-b)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript has a promising conceptual contribution, but the central claim about readout for quantum PDE solvers depends on an equivalence assumption between amplitude-encoded classical solutions and actual quantum solver outputs. This assumption is not tested, and the statistical evidence and missing baselines need strengthening. The issues are addressable in revision, so I recommend major revision rather than rejection. For a quantum information journal, the absence of any run of a quantum PDE solver is a barrier that should be addressed explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First off, the core idea is worth taking seriously: use a QNN to learn a measurement operator that extracts a classification label directly from a quantum state encoding a PDE solution, instead of doing tomography. As a proof of concept for quantum data post-processing, that's a sensible direction, and as far as I can tell the paper is the first to apply it to shock-wave and turbulence detection. The empirical point that basis choice (real vs. Fourier) and coarse-graining strongly affect accuracy is genuinely useful, and the double-QCNN/covariance construction for 2D flow fields is a nice architectural choice. The concept-learning framing is fine, and the paper is clearly written.\n\nBut the central claim—that this addresses the readout problem—is only as strong as an assumption the authors state but do not test. The Results section says the analysis is \"the same as if the quantum data inputs were measured directly from a noiseless quantum PDE solver.\" That equivalence is load-bearing. For the nonlinear PDEs in question, actual quantum solvers typically use Carleman embeddings or Schrodingerisation, producing states in enlarged Hilbert spaces with truncation errors; those states need not look like the amplitude-encoded classical grid functions used here. If the real solver output differs in structure or entanglement, the trained measurement operators may not transfer, and the readout claim falls.\n\nThere are also smaller but real problems. The Burgers experiment has 22 samples; the 100% accuracy is reported after picking the best of three model variants on the full data, with no error bars. There are no classical baselines, so we don't know whether a simple classical classifier on the same amplitude vectors would do equally well or better. Training uses exact statevector expectation values, so shot noise and measurement sampling overhead are never addressed. No code or data are released. None of these are fatal for a proof of concept, but together they mean the paper currently demonstrates a plausible pipeline under idealized conditions, not a solution to the readout problem.\n\nWho gets value: researchers working on quantum algorithms for PDEs or on QML for quantum data. It's a useful pointer for the readout bottleneck, and the basis-choice observation is worth remembering. With a reframing to \"if a solver produces states of this form,\" plus classical baselines and error bars, it would be a solid contribution. As it stands, a serious referee should push for those additions. I'd send it to peer review rather than desk-reject, but I'd want major revision.","headline":"A plausible proof of concept for QNN-based readout of PDE solutions, but the readout claim rests on an untested equivalence to real quantum solver outputs.","tokens_in":14036,"tokens_out":2429,"would_cite":false,"duration_ms":22645,"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":"Quantum PDE readout can be replaced by trained quantum measurements.","keywords":["quantum machine learning","readout problem","quantum PDE solvers","quantum convolutional neural networks","Burgers equation","Navier-Stokes equations","basis choice","quantum data"],"falsifier":"Run the trained Burgers DQNN and Navier-Stokes double-QCNN on states produced by an actual noiseless quantum PDE solver, or by its most faithful simulation, rather than on classically computed amplitude-encoded states; if classification accuracy drops materially below the reported 100% and 92%, the claim that this workflow solves the readout problem for solver output is falsified.","tokens_in":13081,"feed_emoji":"🌊","tokens_out":8552,"duration_ms":76500,"temperature":0.7,"pith_summary":"Quantum differential equation solvers aim to store solutions as $n$-qubit states over $O(2^n)$ grid points, but reading those states out via tomography is exponentially expensive. The paper claims this readout problem can be bypassed when the goal is classification: treat the solver output as quantum data and train a quantum neural network to act as a measurement operator that labels the state directly. It demonstrates the idea on two computational fluid dynamics tasks, classifying Burgers-equation wave solutions as shock or steady and Navier-Stokes cylinder flows as laminar or turbulent. The best Burgers model, a Fourier-basis deep quantum neural network with coarse-graining, reaches 100% accuracy, while the double-QCNN separates laminar from turbulent flow with over 90% accuracy. The paper also shows that the basis in which quantum data is analyzed strongly affects accuracy, making basis choice a central design decision.","feed_headline":"Quantum nets classify shock and turbulent flows without tomography","feed_subtitle":"On quantum states from Burgers and Navier-Stokes solvers, learned observables reach 100% and 92% accuracy.","key_machinery":"The load-bearing mechanism is a trainable measurement operator built from a quantum neural network. A parameterized circuit $U(\\theta)$ acts on the amplitude-encoded PDE solution $\\rho$, and an expectation value of a fixed low-dimensional observable produces the label; for the Burgers problem this is the Pauli-$Z$ observable after a depth-4 DQNN, and for Navier-Stokes it is the covariance $\\mathrm{cov}(O_1,O_2)$ between measurements on two register halves after a double-QCNN. Two supporting tools do essential work: the quantum Fourier transform changes the analysis basis from real space to momentum space with an $O(n^2)$ circuit, and coarse-graining discards part of the register to concentrate on macroscopic features while reducing the qubit count. The basis choice interacts strongly with the architecture, which is why the same model can vary from 27% to 100% accuracy depending on whether Fourier preprocessing and coarse-graining are applied.","core_discovery":"The central claim is that the readout bottleneck of quantum PDE algorithms can be addressed by treating solver outputs as quantum data and learning a hypothesis that is itself a measurement operator. Formally, the solver produces states $\\rho$ in a Hilbert space, and the learned hypothesis $h(\\rho,\\theta)=\\langle\\rho| U^{\\dagger}(\\theta) M U(\\theta)|\\rho\\rangle$ maps each state to a classification label through a parameterized circuit $U(\\theta)$ and a low-dimensional observable $M$. For Burgers shock detection, a deep quantum neural network of depth 4 with Pauli-$Z$ readout achieves 100% accuracy when the data is Fourier-transformed and coarse-grained to half the register; real-space encoding reaches only 27% and Fourier without coarse-graining 45%. For Navier-Stokes flow around a cylinder, a double-QCNN applies separate convolutional networks to the $x$ and $y$ registers and uses the covariance of Pauli-$X$ and Pauli-$Z$ across registers as the hypothesis, reaching 92% average accuracy in real space and 87% in Fourier space with lower variance. The paper reads these results as establishing quantum scientific machine learning as a necessary quantum post-processing step: low-dimensional physical features can be distilled from quantum states without tomographic reconstruction, provided the analysis basis is matched to the problem.","pith_inferences":["If actual noise-bearing quantum PDE solver outputs differ from the amplitude-encoded classical solutions used here in entanglement structure, amplitude distribution, or device noise, the trained measurement operators may need to be retrained on solver-native states; direct transfer is not guaranteed by the paper.","The basis-dependence result suggests a design principle beyond the two examples: choose an analysis basis matched to the symmetries of the target feature, such as momentum space for translationally invariant structures, which could reduce sample complexity in other PDE classification tasks.","The same concept-learning readout could apply to other low-dimensional diagnostics from PDE states, including instability detection, phase-boundary identification, or vortex detection, whenever the desired output is a label rather than the full field.","Combining this readout with shadow-tomography or classical-simulation bounds could reveal when quantum classification is genuinely needed rather than reproducible by classical post-processing of classically computed solutions."],"forward_implications":["Quantum PDE solvers can output physical labels such as shock versus steady or laminar versus turbulent through a small number of measurements instead of full tomography, keeping useful answers accessible as the grid grows.","The analysis basis is a decisive hyperparameter: for Burgers solutions, Fourier basis plus coarse-graining is needed for 100% accuracy, while real-space encoding fails on the same task.","Coarse-graining can reduce the register size and improve generalization, with the Fourier Navier-Stokes model performing better at small training-sample sizes.","Treating solver outputs as quantum data opens a workflow in which quantum machine learning is the readout stage of quantum scientific computing, rather than an optional add-on."],"supporting_citations":[{"why":"Introduces the quantum linear solver and identifies the readout problem that the paper aims to solve.","marker":"[22]"},{"why":"Establishes shadow tomography as a partial readout tool whose cost motivates learned measurement operators.","marker":"[49]"},{"why":"Provides the quantum iterative CFD solver context whose outputs are treated as quantum data.","marker":"[44]"},{"why":"Supplies the quantum-data learning framework that underpins classification directly from states.","marker":"[88]"},{"why":"Supplies the QCNN architecture used in the double-QCNN for Navier-Stokes classification.","marker":"[107]"},{"why":"Supplies the DQNN architecture used for Burgers wave classification.","marker":"[115]"},{"why":"Supplies the quantum Fourier transform used to change the analysis basis before coarse-graining.","marker":"[48]"}],"fun_headline_variants":["Quantum ML reads shock and turbulence from solver states","Learned observables decode quantum PDE outputs sans tomography","Basis-matched quantum learning solves PDE readout bottleneck","Classify shock and turbulence from quantum states without tomography","Quantum nets distill key features from PDE solver states directly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The demonstrations assume that classically computed PDE solutions, amplitude-encoded into quantum states and augmented with white noise, exactly match the states a noiseless quantum differential equation solver would produce; if actual solver states differ in structure, entanglement, or noise, the trained measurement operators may not transfer and the readout advantage would not hold.","fun_headline_variants_meta":{"raw":{"variants":["Quantum ML reads shock and turbulence from solver states","Learned observables decode quantum PDE outputs sans tomography","Basis-matched quantum learning solves PDE readout bottleneck","Classify shock and turbulence from quantum states without tomography","Quantum nets distill key features from PDE solver states directly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1356,"prompt_tokens":982,"completion_tokens":374,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":299}},"tokens_in":598,"tokens_out":374,"duration_ms":4330,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:21:30.336773+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained Burgers DQNN and Navier-Stokes double-QCNN on states produced by an actual noiseless quantum PDE solver, or by its most faithful simulation, rather than on classically computed amplitude-encoded states; if classification accuracy drops materially below the reported 100% and 92%, the claim that this workflow solves the readout problem for solver output is falsified.","supporting_citations":[{"cited_title":"Huang, M","cited_arxiv_id":null,"evidence_quote":"Supplies the DQNN architecture used for Burgers wave classification."}],"review_version":1}