Pith. sign in

REVIEW 2 major objections 5 minor 34 references

Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Mirroring the molecular bond graph in a model's architecture yields competitive molecular property prediction with only 64 trainable parameters, in both quantum and classical implementations, and the matched comparison shows the topology al

desk verdict Useful matched quantum-classical benchmark that cleanly establishes substrate parity, but the abstract overclaims by attributing the efficiency to topology alignment without a non-aligned graph control. read the letter →

arxiv 2607.13737 v2 pith:UNRZ5TEN submitted 2026-07-15 cs.LG

James T. Pegg , Hubert Okadome Valencia , Ronin Wu This is my paper
classification cs.LG MSC 68T0781P68
keywords topology-alignedinductivebiasmolecularpropertypredictionquantummachinelearninggraphneuralnetworksQM9parameterefficiencymatchedbaselinesvariationalcircuits
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that an inductive bias that mirrors the molecular bond graph is the active ingredient behind parameter-efficient learning for molecular property prediction. It instantiates the same bias in two matched architectures—a variational quantum circuit and a classical message-passing network—that share all 64 trainable parameters and identical data splits. Both reach strong test AUCs on QM9 HOMO-LUMO and dipole tasks, saturate within a few hundred training molecules, and keep gradients stable. Because the classical and quantum versions perform alike, the authors conclude that the topology-aligned parameter-sharing scheme, not the quantum substrate, is doing the work, with consequences for how quantum machine learning claims should be benchmarked.

What carries the argument

The central object is the topology-aligned parameter-sharing scheme. Atoms are mapped one-to-one to a fixed register of computational units (qubits in Iso-QGNN, nodes in Iso-CGNN), and bonds determine which pairs interact: the quantum circuit applies a controlled-phase gate plus shared R_Z rotation per bond type, while the classical network passes messages weighted by the same shared bond-type parameters. A single layer, with atom-type embedding (4 params), bond-type embedding (4), per-node update (18), LayerNorm (18), and linear head (20), totals 64 parameters. This shared scheme is what carries the inductive-bias argument: both substrates receive identical inputs, parameter counts, and spl

What would settle it

Run the Iso-CGNN architecture at the same 64-parameter budget and identical QM9 splits but replace the molecular bond edges with random edges (or a fully connected graph) drawn from the same degree distribution; if test AUC stays at the reported ~0.9 gap / ~0.78 dipole levels rather than falling toward the 0.66–0.70 logistic-regression control, the bond-graph alignment is not the active ingredient claimed.

Watch

Extended reading notes

Core claim

Following the paper: the central claim is that topology-aligned architectures—where atoms occupy a fixed register of computational units and bonds select which pairs interact via shared per-bond-type parameters—achieve competitive binary classification of HOMO-LUMO gap and dipole moment on QM9 with only 64 trainable parameters in both a quantum circuit (Iso-QGNN) and a parameter-matched classical network (Iso-CGNN). On identical splits the quantum and classical implementations perform comparably (gap AUC about 0.89 and 0.92, dipole about 0.77–0.79), both far above a five-parameter composition-only logistic regression control. The paper further claims these models reach 90% of asymptotic perf

Load-bearing premise

The paper attributes the observed parameter efficiency specifically to topology alignment, but its only non-graph control is a composition-only logistic regression; without a matched-parameter graph model that lacks bond-topology alignment (e.g., random or full connectivity), the gain could in principle stem from any graph-structured inductive bias.

Editorial extensions

If this is right

  • With 64 trainable parameters, both implementations reach test AUCs around 0.89–0.92 on the HOMO-LUMO gap and ~0.78 on the dipole moment, far outperforming a composition-only logistic regression control.
  • Both architectures reach 90% of their asymptotic test AUC within roughly 100–300 training molecules, indicating strong data efficiency in the low-data regime common in quantum chemistry.
  • Gradient norms remain of order 1e-1 to 1e0 throughout training for both substrates at QM9 scale, so the topology-aligned sharing scheme does not exhibit barren-plateau-like vanishing gradients in this setting.
  • Because the quantum and classical implementations are statistically indistinguishable (or only slightly different) across tasks, matched-baseline controls should be considered necessary methodology when evaluating quantum ML models for small molecular property tasks.
  • The qualitative ordering in which the classical implementation saturates faster on the gap task suggests future work could examine where (if anywhere) the quantum substrate provides an advantage—for example with non-diagonal observables or larger systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive test of the paper's interpretive claim: train the same 64-parameter architecture on the same splits but with randomized or fully connected edges instead of the molecular bond graph; if the gain over the logistic-regression control persists, the active ingredient is generic graph structure, not bond topology specifically.
  • The paper's own bond-parameter analysis suggests chemically informed initialization (e.g., scaling bond weights by bond enthalpies or related descriptors) could make the topology alignment stronger and yield more stable optimization, which is a natural testable extension.
  • The dipole task bottleneck is attributed to missing 3D geometry; an equivariant extension that injects coordinates should recover dipole performance, and the matched-comparison framework could quantify that gain.
  • The framework's single-layer restriction will likely need more layers as molecular size grows; the constant-in-N shared embeddings suggest the parameter-count advantage may persist, but this is untested beyond QM9's 9 heavy atoms.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The manuscript proposes a topology-aligned inductive bias for molecular property prediction, in which the architecture mirrors the molecular bond graph: atoms map to a fixed register of computational units and bonds determine which pairs interact through shared learnable parameters. This bias is instantiated in a variational quantum circuit (Iso-QGNN) and a parameter-matched classical message-passing network (Iso-CGNN), each with 64 trainable parameters. On QM9 binary classification of HOMO-LUMO gap and dipole moment, the models reach test AUCs of about 0.89 (quantum) and 0.92 (classical) on the gap task, and about 0.78 on the dipole task, well above a composition-only logistic regression control. The paper further reports data efficiency (90% of asymptotic performance within about 300 training molecules) and stable gradient norms, and argues that the topology-aligned parameter-sharing scheme—not the quantum substrate—is the active ingredient. The authors release code for all experiments and provide an explicit parameter accounting.

Significance. If the central claim is established, the paper makes a useful methodological contribution to QML benchmarking: matched classical controls can isolate the inductive bias from the quantum substrate, and the proposed topology-aligned architecture is a clean testbed. The manuscript is commendably transparent: it includes an explicit parameter table, matched splits across architectures, honest discussion of seed sensitivity, a public code repository reproducing all figures and tables, and candid limitations about barren-plateau scaling and the dipole-task geometric bottleneck. These strengths make the paper potentially valuable for the QCE community. However, the strongest interpretive claim—that topology alignment, not graph structure in general, is the active ingredient—is not currently supported by the experimental design.

major comments (2)
  1. [Section III-A and III-E; Table III] The central claim that the 64-parameter compactness, data efficiency, and stable gradients are properties of the 'topology-aligned parameter-sharing scheme' is underdetermined by the experiments. The only non-graph control is a five-parameter logistic regression on heavy-atom composition; this controls for composition but not for graph structure. To attribute the observed gains specifically to mirroring the molecular bond graph, the authors must run a matched-capacity graph model with the same 64-parameter shared scheme but a non-topology-aligned connectivity, e.g., a random graph or fully connected graph with the same number of interactions per node. If such a control performs equally well, the active ingredient would be generic graph-structured message passing with shared parameters, not bond-graph alignment. This is a concrete, load-bearing experiment for the abstract and Section III-
  2. [Section III-B; Fig. 3] The sample-complexity sweep uses a different training protocol from the matched-capacity experiments: 40 epochs, validation-selected reporting, and a separate 3000-molecule pool with a fixed 500-molecule test set. The saturation point N_90% is defined as the first training size at which the model reaches 90% of its own asymptotic peak AUC, but if the 40-epoch budget is insufficient for convergence, the 'asymptotic peak' is not well defined and the reported N_90% values may be artifacts of the protocol. Since data efficiency is one of the paper's three headline findings, the authors should either verify that the sweep results are robust to a longer training budget for at least several training sizes, or present the sweep explicitly as an exploratory illustration rather than a quantitative claim.
minor comments (5)
  1. [Table III] The 'Params' column entries such as '640.887' appear to concatenate the parameter count and the AUC value; these should be separate columns (e.g., '64' and '0.887'). The same applies to '50.659' and '50.698'.
  2. [Fig. 3 caption and Section III-B] The caption says the held-out set is a fixed 500-molecule set, while the text says the sweep uses a 3000-molecule pool with a 500-molecule held-out test set. Clarify the relationship between the training pool and the held-out set, and state the validation split used for epoch selection.
  3. [Section II-C] The sentence 'the variational quantum models show sensitivity to initialisation ... and shorter training runs occasionally leave a seed under-converged' is a useful caveat, but it raises a question about the 200-epoch convergence criterion. State how convergence was checked (e.g., loss curves reaching a plateau) so readers can assess whether the reported means are true converged values.
  4. [Section III-D] The discussion of bond-parameter non-robustness is honest and well framed. The phrase 'the underlying weight distributions remain task- and substrate-dependent' is slightly vague; consider reporting a quantitative distributional distance (e.g., KL divergence or Wasserstein distance) to support the claim that the joint distributions differ across tasks and substrates.
  5. [Section III-F] The scaling analysis is reasonable, but the statement 'for organic molecules, d_max is bounded by carbon valence' is only true for standard valence models; some QM9 molecules contain hypervalent or resonance structures. A more precise statement would be 'bounded by a small constant for the QM9 chemical space'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an empirical matched-capacity comparison; the central interpretive claim is underdetermined by experimental design, not derived from its own assumptions.

full rationale

The paper's derivation chain is empirical rather than definitional. The topology-aligned architectures are constructed explicitly, and the reported test AUCs, sample-complexity curves, and gradient norms are measured outcomes, not quantities fitted to the conclusions they support. The N90% saturation metric is defined as each model reaching 90% of its own asymptotic peak AUC; this is descriptively self-referential but does not constitute a prediction or a fitted input called a prediction, and the paper presents it as a descriptive saturation point rather than as validation of the topology-alignment hypothesis. The central interpretive claim — that the topology-aligned inductive bias, rather than the quantum substrate, drives parameter efficiency — is supported by comparing two parameter-matched architectures that share the same alignment and differ only in substrate, plus a non-graph logistic regression control. The missing control is a matched-capacity graph model without bond-topology alignment (e.g., random or full connectivity), so the attribution to the specific bond-graph alignment is underdetermined. However, this is an experimental identification gap, not circularity: no equation reduces to an assumed result, and no fitted parameter is renamed as a prediction. The only overlapping-author citation (ref. [4], Qca-MolGAN) is contextual background in the introduction and is not load-bearing for any claim. No uniqueness theorem, ansatz-smuggling via citation, or renaming of a known result occurs. The paper is self-contained as an empirical benchmark, and the circularity burden is accordingly low.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central claim is empirical; the only fitted quantities are the model's 64 trainable parameters, plus per-trial median thresholds used to build binary labels. The main assumed premises are domain assumptions about QM9, the median-split task framing, the faithfulness of the classical analogue, and the representativeness of the 1500-molecule sample. No new physical entities are introduced.

free parameters (5)
  • atom-type embeddings (4) = learned
    Per-atom-type rotation angles θ_atom for C,N,O,F; trained on data.
  • bond-type embeddings (4) = learned
    Per-bond-type interaction strengths θ_bond for single/double/triple/aromatic; trained on data.
  • per-node update parameters (18) = learned
    Two parameters per node (QGNN: φ1,φ2; CGNN: s,b) for the rotation/node-update layer.
  • readout LayerNorm parameters (18) = learned
    γ and β for each of the 9 node features.
  • linear classifier weights and bias (20) = learned
    Output layer mapping 9 node features to 2 logits.
assumptions (7)
  • domain assumption QM9 reference labels (HOMO-LUMO gap, dipole moment) are accurate enough for benchmarking.
    The paper uses QM9 as ground truth without question.
  • domain assumption Binary classification via median split on the training set preserves the signal relevant to parameter efficiency.
    Median-split labels are used throughout; the paper does not validate against regression performance.
  • domain assumption Heavy-atom graph with implicit hydrogens retains sufficient chemical information for these tasks.
    The graph representation drops hydrogen atoms and spatial geometry; the paper acknowledges the dipole task bottleneck.
  • domain assumption Ghost-node padding gives a size signal that is identical across architectures and no unfair advantage.
    Padding is asserted to confer no quantum-versus-classical advantage, but no explicit test is performed.
  • ad hoc to paper The classical analogue in Table I is a faithful parameter-matched counterpart to the quantum circuit.
    The claim that the comparison isolates substrate from architecture depends on this mapping; CZ+RZ and message passing with tanh are not exactly equivalent.
  • domain assumption State-vector simulation faithfully represents the noiseless quantum circuit.
    PennyLane simulation is used; no hardware noise model is included.
  • standard math Adam optimization with the stated hyperparameters converges to a representative model.
    Standard ML assumption; convergence is checked by 200 epochs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction." pith.science (2026). https://pith.science/paper/UNRZ5TEN

@misc{pith2026260713737,
  author       = {Pith},
  title        = {Pith review of: Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UNRZ5TEN}},
  note         = {Machine review of arXiv:2607.13737}
}
read the original abstract

For low-data and resource-constrained regimes typical of quantum chemistry, parameter-efficient learning is a key objective. Here, we propose a topology-aligned inductive bias in which the model architecture mirrors the molecular bond graph: atoms map to a fixed register of computational units, and bonds determine which pairs interact through shared learnable parameters. This principle is instantiated in two architectures: a variational quantum circuit (Iso-QGNN) and a parameter-matched classical message-passing network (Iso-CGNN). The models are benchmarked on HOMO-LUMO and dipole moment binary classification tasks over the QM9 benchmark. With 64 trainable parameters, the implementations achieve test AUCs of approximately 0.89 (quantum) and 0.92 (classical) on the gap task, and close to 0.78 (both) on the dipole task. The models reach 90% of asymptotic performance within about 300 training molecules and gradient norms remain stable throughout training. These results indicate that the topology-aligned inductive bias is the active ingredient driving parameter efficiency at QM9 scale, with implications for matched-baseline benchmarking in quantum machine learning.

Figures

Figures reproduced from arXiv: 2607.13737 by the authors.

Figure 1
Figure 1. Iso-QGNN circuit for a three-atom example (for￾mamide O=CH–NH2, with implicit H atoms). Each atom is assigned a qubit initialised in |0⟩ and rotated by RY (θatom) with a per-atom-type parameter. Bonded atoms are then entan￾gled with a CZ gate followed by symmetric RZ(θbond) rota￾tions with per-bond-type parameters (single and double bonds receive different parameters). A final RX, RY rotation layer is applied, and e… view at source ↗
Figure 2
Figure 2. Confusion matrices. Aggregated test-set confusion over ten trials. Rows are the architectures (Iso-QGNN upper, Iso-CGNN lower); columns are the tasks (HOMO-LUMO gap left, dipole moment right). Diagonal dominance confirms both architectures learn discriminative features rather than class priors; Iso-CGNN’s stronger diagonal on the gap task appears in the lower-left panel. The architectures encode purely two-dimension… view at source ↗
Figure 3
Figure 3. Sample complexity. Test AUC on a fixed 500- molecule held-out set as a function of training-set size. The architectures saturate within N ≤ 300 molecules. Iso-CGNN reaches saturation faster on both tasks and maintains a small but consistent lead on the gap task across all training sizes. Error bars are standard deviations across five trials. For each training-set size we report the test AUC at the epoch selected by … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Gradient flow stability. Mean gradient norms (log scale) per parameter group across training on the gap task, for Iso-QGNN (upper graph) and Iso-CGNN (lower graph). Both maintain non-vanishing gradients (of order 10−1 to ∼ 100 ); the quantum implementation (upper) clim…
Figure 5
Figure 5. Figure 5: Learned bond parameters. Mean and standard de￾viation of the bond-type weights across ten trials, for both architectures on the HOMO-LUMO gap (upper graph) and dipole moment (lower graph) tasks. The large error bars indicate that neither architecture converges to a sta…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 5 linked inside Pith

  1. [1]

    E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials,

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, “E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials,” Nature Communications, vol. 13, no. 1, p. 2453, 2022

    2022

  2. [2]

    Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations,

    G. Wang, C. Wang, X. Zhang, Z. Li, J. Zhou, and Z. Sun, “Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations,”iScience, vol. 27, no. 5, p. 109673, 2024

    2024

  3. [3]

    Molgan: An implicit generative model for small molecular graphs,

    N. De Cao and T. Kipf, “Molgan: An implicit generative model for small molecular graphs,”arXiv preprint arXiv:1805.11973, 2018

    arXiv 2018

  4. [4]

    Qca-molgan: Quantum circuit associative molecular gan with multi- agent reinforcement learning,

    A. M. Thomas, Y .-C. Chen, H. O. Valencia, S. T. Jose, and R. Wu, “Qca-molgan: Quantum circuit associative molecular gan with multi- agent reinforcement learning,”arXiv preprint arXiv:2509.05051, 2025

    arXiv 2025

  5. [5]

    Hybrid quantum-classical machine learning for generative chemistry and drug design,

    S. M. Mniszewski, C. F. A. Negre, H. Ushijima-Mwesigwaet al., “Hybrid quantum-classical machine learning for generative chemistry and drug design,”Scientific Reports, vol. 11, p. 4048, 2021

    2021

  6. [6]

    Coupled-cluster theory in quantum chemistry,

    R. J. Bartlett and M. Musiał, “Coupled-cluster theory in quantum chemistry,”Reviews of Modern Physics, vol. 79, no. 1, pp. 291–352, 2007

    2007

  7. [7]

    Simulating physics with computers,

    R. P. Feynman, “Simulating physics with computers,”International Journal of Theoretical Physics, vol. 21, no. 6–7, pp. 467–488, 1982

    1982

  8. [8]

    Quantum chemistry in the age of quantum computing,

    Y . Cao, J. Romero, J. P. Olson, M. Degroote, P. D. Johnson, M. Kiefer- ová, I. D. Kivlichan, T. Menke, B. Peropadre, N. P. D. Sawayaet al., “Quantum chemistry in the age of quantum computing,”Chemical Reviews, vol. 119, no. 19, pp. 10 856–10 915, 2019

    2019

Show all 34 references
  1. [9]

    Machine learning for molecular and materials science,

    K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, “Machine learning for molecular and materials science,”Nature, vol. 559, no. 7715, pp. 547–555, 2018

    2018

  2. [10]

    Ani-1: an extensible neural network potential with dft accuracy at force field computational cost,

    J. S. Smith, O. Isayev, and A. E. Roitberg, “Ani-1: an extensible neural network potential with dft accuracy at force field computational cost,” Chemical Science, vol. 8, no. 4, pp. 3192–3203, 2017

    2017

  3. [11]

    Benchmarking graph neural networks for materials chemistry,

    V . Fung, J. Zhang, E. Juarez, and B. G. Sumpter, “Benchmarking graph neural networks for materials chemistry,”npj Computational Materials, vol. 7, no. 1, p. 84, 2021

    2021

  4. [12]

    Supervised learning with quantum- enhanced feature spaces,

    V . Havlí ˇcek, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, “Supervised learning with quantum- enhanced feature spaces,”Nature, vol. 567, no. 7747, pp. 209–212, 2019

    2019

  5. [13]

    The power of quantum neural networks,

    A. Abbas, D. Sutter, C. Zoufal, A. Lucchi, A. Figalli, and S. Woerner, “The power of quantum neural networks,”Nature Computational Sci- ence, vol. 1, no. 6, pp. 403–409, 2021

    2021

  6. [14]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veliˇckovi´c, “Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,”arXiv preprint arXiv:2104.13478, 2021

    2021 arXiv

  7. [15]

    E(n) equivariant graph neural networks,

    V . G. Satorras, E. Hoogeboom, and M. Welling, “E(n) equivariant graph neural networks,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 9323–9332

    2021

  8. [16]

    Equivariant diffusion for molecule generation in 3d,

    E. Hoogeboom, V . G. Satorras, C. Vignac, and M. Welling, “Equivariant diffusion for molecule generation in 3d,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 8867–8887

    2022

  9. [17]

    Better than classical? The subtle art of benchmarking quantum machine learning models,

    J. Bowles, S. Ahmed, and M. Schuld, “Better than classical? The subtle art of benchmarking quantum machine learning models,”Quantum, vol. 8, p. 1559, 2024

    2024

  10. [18]

    Quantum kernel methods under scrutiny: a benchmarking study,

    J. Egginger, A. Sakhnenko, and J. M. Lorenz, “Quantum kernel methods under scrutiny: a benchmarking study,”Quantum Machine Intelligence, vol. 7, p. 47, 2025

    2025

  11. [19]

    Benchmarking quantum machine learning kernel training for classification tasks,

    D. Alvarez-Estevez, “Benchmarking quantum machine learning kernel training for classification tasks,”arXiv preprint arXiv:2408.10274, 2024

    2024 arXiv

  12. [20]

    Quantum computing in the NISQ era and beyond,

    J. Preskill, “Quantum computing in the NISQ era and beyond,”Quan- tum, vol. 2, p. 79, 2018

    2018

  13. [21]

    Solving the graph-isomorphism problem with a quantum annealer,

    I. Hen and A. P. Young, “Solving the graph-isomorphism problem with a quantum annealer,”Physical Review A, vol. 86, no. 4, p. 042310, 2012

    2012

  14. [22]

    Enhancing quantum annealing per- formance for the molecular similarity problem,

    M. Hernandez and M. Aramon, “Enhancing quantum annealing per- formance for the molecular similarity problem,”Quantum Information Processing, vol. 16, no. 5, p. 133, 2017

    2017

  15. [23]

    Solving quantum chemistry problems with a D-Wave quantum annealer,

    M. Streif, F. Neukart, and M. Leib, “Solving quantum chemistry problems with a D-Wave quantum annealer,” inInternational Workshop on Quantum Technology and Optimization Problems. Springer, 2019, pp. 111–122

    2019

  16. [24]

    Scalable quantum simulation of molecular energies,

    P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero, J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tranter, N. Dinget al., “Scalable quantum simulation of molecular energies,”Physical Review X, vol. 6, no. 3, p. 031007, 2016

    2016

  17. [25]

    Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets,

    A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, “Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets,”Nature, vol. 549, no. 7671, pp. 242–246, 2017

    2017

  18. [26]

    Noisy intermediate-scale quantum algorithms,

    K. Bharti, A. Cervera-Lierta, T. H. Kyaw, T. Haug, S. Alperin-Lea, A. Anand, M. Degroote, H. Heimonen, J. S. Kottmann, T. Menkeet al., “Noisy intermediate-scale quantum algorithms,”Reviews of Modern Physics, vol. 94, no. 1, p. 015004, 2022

    2022

  19. [27]

    Efficient symmetry-preserving state preparation circuits for the variational quantum eigensolver algorithm,

    B. T. Gard, L. Zhu, G. S. Barron, N. J. Mayhall, S. E. Economou, and E. Barnes, “Efficient symmetry-preserving state preparation circuits for the variational quantum eigensolver algorithm,”npj Quantum Informa- tion, vol. 6, no. 1, p. 10, 2020

    2020

  20. [28]

    Barren plateaus in quantum neural network training landscapes,

    J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, “Barren plateaus in quantum neural network training landscapes,”Nature Communications, vol. 9, no. 1, p. 4812, 2018

    2018

  21. [29]

    Cost function dependent barren plateaus in shallow parameterized quantum circuits,

    M. Cerezo, A. Sone, T. V olkoff, L. Cincio, and P. J. Coles, “Cost function dependent barren plateaus in shallow parameterized quantum circuits,” Nature Communications, vol. 12, no. 1, p. 1791, 2021

    2021

  22. [30]

    Deep variational quantum eigensolver: A divide-and-conquer method for solving a larger problem with smaller size quantum com- puters,

    K. Fujii, K. Mizuta, H. Ueda, K. Mitarai, W. Mizukami, and Y . O. Nak- agawa, “Deep variational quantum eigensolver: A divide-and-conquer method for solving a larger problem with smaller size quantum com- puters,”PRX Quantum, vol. 3, no. 1, p. 010346, 2022

    2022

  23. [31]

    Quantum graph neural network models for materials search,

    J.-Y . Ryu, E. Elala, and J.-K. K. Rhee, “Quantum graph neural network models for materials search,”Materials, vol. 16, no. 12, p. 4300, 2023

    2023

  24. [32]

    Quantum chemistry structures and properties of 134 kilo molecules,

    R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. von Lilienfeld, “Quantum chemistry structures and properties of 134 kilo molecules,” Scientific Data, vol. 1, no. 1, p. 140022, 2014

    2014

  25. [33]

    Pennylane: Automatic differentiation of hybrid quantum-classical computations,

    V . Bergholm, J. Izaac, M. Schuld, C. Gogolin, M. S. Alam, S. Ahmed, J. M. Arrazola, C. Blank, A. Delgado, S. Jahangiriet al., “Pennylane: Automatic differentiation of hybrid quantum-classical computations,” arXiv preprint arXiv:1811.04968, 2018

    2018 arXiv

  26. [34]

    Trapped- ion quantum computing: Progress and challenges,

    C. D. Bruzewicz, J. Chiaverini, R. McConnell, and J. M. Sage, “Trapped- ion quantum computing: Progress and challenges,”Applied Physics Reviews, vol. 6, no. 2, p. 021314, 2019

    2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.