Pith. sign in

REVIEW 5 major objections 5 minor 45 references

Hybrid Quantum-Classical Inverse Design of Metasurfaces for Tailored Narrow Band Absorption

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Hybrid quantum-classical GAN claims 10x faster metasurface design.

desk verdict A plausible application of an existing QGAN architecture to metasurface inverse design, undermined by internally inconsistent efficiency claims and no controlled baseline comparison. read the letter →

arxiv 2507.18127 v2 pith:KVWBPHB5 submitted 2025-07-24 physics.optics cond-mat.soft

Sreeraj Rajan Warrier , Jayasri Dontabhaktuni This is my paper
classification physics.opticscond-mat.soft
keywords metasurfaceinversedesignquantumgenerativeadversarialnetworkvariationalautoencodernarrow-bandabsorptionFanoresonanceQ-factorlatentspacelearninghybridquantum-classicalcomputing
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 attempts to establish that a hybrid quantum-classical generative model, the Latent Style-based Quantum GAN (LaSt-QGAN), can inverse-design metasurfaces for narrow-band absorption far more efficiently than a classical GAN. The model joins a pretrained variational autoencoder with a quantum generator and a classical discriminator, and it learns to map target absorption spectra to metasurface unit-cell layouts that are encoded as RGB images. The authors report a tenfold reduction in training time (2.5 hours versus 25), a fortyfold reduction in training data (500 versus 20,000 samples), and a mean squared error one order of magnitude lower than the classical baseline's. They further report that the trained model generates unit cells with Q-factors around $10^4$, even though the training data only contains Q-factors around $10^3$. If these claims hold, quantum-enhanced machine learning would supply a practical route to cheaper and sharper metasurface design.

What carries the argument

The load-bearing mechanism is a latent-space quantum GAN built around a variational autoencoder. The autoencoder (either a $\beta$-VAE or an importance-weighted autoencoder) compresses $64 \times 64 \times 3$ RGB-encoded metasurface images—red channel for plasma frequency, green for refractive index, blue for dielectric thickness—into a low-dimensional latent vector. The quantum generator is a style-based variational circuit in which the concatenation of latent noise $z$ and a conditional vector $\gamma$ (four Fano parameters fitted to the target absorption spectrum) is mapped by an affine transform into rotation angles for $RY$ gates, interlaced with circular CNOT entanglement, and read out through Pauli-Z expectation values. A classical fully connected discriminator distinguishes real from generated latent codes. This compression is what lets a small NISQ-scale circuit handle a design problem that would otherwise require pixel-space generation.

What would settle it

Train the classical GAN of Yeung et al. on the same 500-design subset, with the same Fano-parameter conditioning, data split, and GPU hardware, and compare its runtime and MSE with LaSt-QGAN's; if the classical GAN reaches an MSE near $10^{-4}$ in about 2.5 hours, the reported quantum advantage disappears.

Watch

Extended reading notes

Core claim

The central claim is that LaSt-QGAN outperforms the classical GAN of Yeung et al. across all reported performance metrics: training time falls from 25 hours to 2.5 hours, the mean squared error falls from $10^{-3}$ to $10^{-4}$, and the required training set falls from 20,000 to 500 distinct metasurface layouts. The paper also claims that the quantum generator produces asymmetric, freeform unit-cell geometries whose Fano resonances reach Q-factors of order $10^4$, while the training data only reaches order $10^3$; this extrapolation is attributed to symmetry breaking that suppresses radiative losses. A material look-up table is used to replace predicted optical constants with conventional materials, and the re-simulated designs retain about 95% precision relative to the original predictions.

Load-bearing premise

The reported 10x and 40x improvements assume that the 500-design subset of the Yeung et al. database and the published GAN baseline are directly comparable, using identical design targets, data splits, and hardware; the paper does not state the subset sampling rule or show that the baseline was re-run under the same conditions.

Editorial extensions

If this is right

  • If the reported metrics hold, inverse design of absorbing metasurfaces would need 40x fewer training samples, cutting the costly electromagnetic simulation burden for dataset construction.
  • The 10x reduction in training time would make iterative design loops practical on a single GPU rather than a large compute cluster.
  • Generating Q-factors beyond the training range suggests the model can explore high-Q design regions whose sharp resonances are useful for sensing, filtering, and thermal emission.
  • Substituting predicted materials with near-matching conventional ones, while retaining 95% spectral precision, brings the generated designs one step closer to fabrication.

Reading between the lines

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

  • The comparison is made against a published classical baseline, not against a classical GAN re-trained on the same 500-sample subset under identical conditions; rerunning the baseline this way would isolate whether the speedup comes from the quantum generator or from the smaller, curated dataset.
  • Because the conditional vector uses only four Fano parameters, the same architecture could plausibly be retargeted to other spectral line shapes, such as Lorentzian absorbers or multi-peak filters, by swapping the conditioning features.
  • The high-Q extrapolation is a simulation-level claim; fabricating and measuring one of the generated freeform designs would test whether real devices reach Q-factors near $10^4$ or stay at the level of the training data.
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

5 major / 5 minor

Summary. The paper proposes LaSt-QGAN, a hybrid quantum-classical generative model that combines a variational autoencoder (VAE) with a quantum GAN for inverse design of metasurfaces with tailored narrow-band absorption. The method is tested on MIM and hybrid dielectric metasurface unit cells encoded as RGB images, with a limited dataset of 500 layouts drawn from the multiclass dataset of Yeung et al. [14,31]. The central claims are a 10x reduction in training time, a 40x reduction in the number of training samples, and an order-of-magnitude lower MSE compared with the classical GAN baseline of [14], plus extrapolation to Q-factors of order 10^4 from training data with Q-factors of order 10^3. Validation is performed with COMSOL simulations of the generated designs, and a material look-up table is used to substitute predicted material properties with real alternatives.

Significance. If the claimed improvements were established by controlled comparison, the paper would be significant: it would demonstrate that a hybrid quantum-classical generator can match or exceed a classical GAN in a practical inverse-design task while using far fewer training samples, and it would provide a concrete example of a quantum model extrapolating beyond its training distribution. The use of an external simulation tool (COMSOL) to validate generated designs is a strength, as is the material substitution step for manufacturability. However, the quantitative claims are currently not well supported because the baseline is not re-run under matched conditions and the reported metrics lack statistical grounding, so the significance for quantum advantage in photonics remains unproven.

major comments (5)
  1. [Conclusions vs. Table 1] The stated training-time improvements are internally inconsistent. Table 1 reports 2.5 h versus 25 h (a 10x reduction), and the abstract and first paragraph of the Conclusions state a 10x/'one-tenth' reduction, but the final sentence of Sec. 4 says 'achieving a 40% reduction in training time without compromising accuracy.' A 40% reduction (1.67x) cannot be reconciled with a 10x reduction. This is a load-bearing inconsistency because the headline claim of the paper is the computational speedup.
  2. [Table 1 and Sec. 2 (dataset)] The comparison in Table 1 is not controlled. The GAN baseline metrics (25 h, 10^-3 MSE, 20,000 samples) are taken directly from [14], but the present work uses only 500 layouts from seven shape types drawn from [14,31], with no description of the sampling rule. No evidence is provided that the classical GAN was re-trained on the same 500-design subset, the same conditional targets, the same train/validation split, or the same DGX-1 hardware. Without such a matched comparison, the 10x/40x improvements cannot be attributed to the hybrid quantum-classical method; they may simply reflect the smaller dataset or a faster GPU.
  3. [Sec. 2 (training-time accounting)] The reported 2.5 h training time does not account for all components of the pipeline. The quantum generator is executed on a PennyLane simulator running on an A100 GPU, so the measured time includes classical simulation overhead of the quantum circuits. In addition, the VAE is pre-trained on the image dataset, and that pretraining cost is not reported. It is therefore unclear whether 'training time' refers only to the QGAN stage or to the full LaSt-QGAN pipeline, and whether the 25 h figure from [14] includes comparable pretraining or data-generation steps.
  4. [Table 1 and Sec. 3 (statistical support)] All performance metrics (MSE 10^-4 versus 10^-3, runtime 2.5 h versus 25 h) are point values without error bars, confidence intervals, or multiple seeded runs. GAN and QGAN training are stochastic, and single-run comparisons are insufficient to support the claim that LaSt-QGAN 'significantly outperforms' the classical GAN across all key metrics. At a minimum, repeated runs with different random seeds should be reported.
  5. [Sec. 3 and Fig. 9 (Q-factor extrapolation)] The claim that the model generates Q-factors of order 10^4 while the training data has Q-factors up to 10^3 is supported by only four examples in Fig. 9, with no error bars or sensitivity analysis. Since high-Q resonances are sensitive to geometric tolerances, it would be useful to know how robust these generated designs are to small perturbations in the predicted geometries and whether the four examples are representative or exceptional. This is a secondary but relevant part of the abstract's central claim.
minor comments (5)
  1. [Throughout] The model name is inconsistent: 'LaSt-QGAN' in the abstract, 'La-St QGAN' and 'LaSt-QGAN' in Sec. 3. Please standardize.
  2. [Sec. 2] The DGX-1 description states 'a single A100 (Volta) GPU'; the NVIDIA A100 is based on the Ampere architecture, not Volta. Please correct.
  3. [Eq. (1)] The notation 'Ex− →pdata' is typeset incorrectly; the arrow and subscript should be formatted as a standard expectation over the data distribution.
  4. [Sec. 2 (dataset description)] The sentence describing the dataset says 'structures presented in this study were derived from seven unique shape types, as outlined in [14,31]' but does not state the sampling rule (e.g., random, stratified, or hand-picked). This is relevant to the reproducibility of the 500-design subset.
  5. [Sec. 3 (material substitution)] The '95% precision' claim for the material look-up table is not defined; please state the metric used (e.g., spectral overlap, normalized RMSE) and how it is computed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the inverse-design outputs are verified by external COMSOL simulation, and the high-Q claim is a conditional-target realization rather than a reduction to training inputs.

full rationale

The central derivation chain is empirical rather than definitional. Metasurface layouts are RGB-encoded, projected into a VAE latent space, and a QGAN generator conditioned on Fano-fitted target spectra outputs unit-cell geometries. Those geometries are then independently re-simulated in COMSOL, and the simulated absorption spectra are compared with the target spectra. This validation is external to the GAN training loss, so the claimed 'prediction' of absorption responses is not equivalent to the training inputs by construction. The high-Q claim is the closest thing to a circular-looking statement: Q-factor is related to the fitted Fano parameters (resonance frequency and linewidth) that form the conditional vector, so saying the model 'generates Q-factor up to 10^4' means it can realize out-of-distribution high-Q target spectra, not that it discovers them without being told. Since the resulting geometries are still verified by COMSOL simulation, this is a legitimate conditional inverse-design test rather than a reduction. The architecture is credited to an external prior work [25], not to the current authors, and no load-bearing self-citation or uniqueness theorem is invoked. Table 1's comparison with Yeung et al. [14] is a fairness concern because the baseline was not re-run on the same 500-layout subset, and the Conclusions contain an internal inconsistency ('reduces training time by one-tenth' versus '40% reduction in training time'), but these are reporting and benchmarking weaknesses, not circular reasoning. Accordingly, no circular step meets the evidentiary bar of being equivalent to its input by definition.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on domain assumptions about the Fano parameterization, COMSOL simulation fidelity, the representativeness of the 500-sample subset, the sufficiency of the VAE latent space, and the fairness of the baseline comparison. No new physical particles, forces, or dimensions are introduced, and the predicted material parameters are just numerical outputs substituted with real materials from a lookup table.

free parameters (4)
  • beta (VAE regularization weight) = 2
    Set to 2 for both the beta-VAE and IWAE; no sensitivity analysis or justification is provided, and it controls the structure of the latent space used by the QGAN.
  • IWAE importance samples = 5
    Fixed at 5 for the importance weighted autoencoder; the choice is not varied or justified.
  • quantum generator initialization range = [-0.01, 0.01]
    Orthogonal weight matrices and biases are initialized from this uniform range to mitigate gradient problems, but the range is a hand-set hyperparameter.
  • number of qubits and layers = not stated
    The quantum circuit uses RY rotations and circular CNOT entanglement, but the actual number of qubits and layers is never given, leaving model capacity as an uncontrolled free choice.
assumptions (5)
  • domain assumption Fano equation adequately represents the absorption spectra used for conditioning
    Section 2 states that resonance frequency, linewidth, peak absorption, and Fano asymmetry are derived by fitting and used as the conditional vector; if the fit is lossy, the inverse mapping is incomplete.
  • domain assumption COMSOL simulations accurately model the fabricated metasurface response
    Training spectra and validation of generated designs come from COMSOL Multiphysics simulations in Sections 2 and 3, with no experimental measurements.
  • domain assumption The 500-sample subset of the Yeung et al. dataset is representative of the full design space
    Section 2 says the training dataset consists of 500 layouts from seven shape types, but the sampling rule is unexplained, and all efficiency and accuracy claims depend on this subset.
  • domain assumption The VAE latent space preserves the information needed for inverse design
    Section 2 pretrains a VAE to compress 64x64x3 images into latent features; if reconstruction is lossy, decoded QGAN outputs may not correspond to physically valid designs.
  • domain assumption The classical GAN baseline in [14] is comparable to the proposed model
    Table 1 compares runtime, MSE, and sample count against Yeung et al. without specifying identical splits, targets, preprocessing, or hardware, so the comparison may not be fair.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Quantum-Classical Inverse Design of Metasurfaces for Tailored Narrow Band Absorption." pith.science (2026). https://pith.science/paper/KVWBPHB5

@misc{pith2026250718127,
  author       = {Pith},
  title        = {Pith review of: Hybrid Quantum-Classical Inverse Design of Metasurfaces for Tailored Narrow Band Absorption},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KVWBPHB5}},
  note         = {Machine review of arXiv:2507.18127}
}
abstract

The inverse design of metasurfaces poses a considerable challenge because of the intricate interdependencies that exist between structural characteristics and electromagnetic responses. Traditional optimization methods require significant computational resources and frequently do not produce the most effective solutions. This study presents a hybrid quantum-classical machine learning approach known as Latent Style-based Quantum GAN (LaSt-QGAN). This method integrates a Variational Autoencoder (VAE) with a Quantum Generative Adversarial Network (QGAN) to enhance the optimization of metasurface designs aimed at achieving narrow-band absorption and unidirectionality. The proposed method results in a reduction of training time by 10X and a decrease in data requirements by 40X when compared to traditional GAN-based approaches. The produced metasurface designs demonstrate a high fidelity in relation to the target absorption spectra compared to the classical GAN based methods. Additionally, the integration of a material look-up table facilitates manufacturability by allowing for the substitution of predicted material properties with viable alternatives, all while preserving performance accuracy. Moreover the model is able to generate Q-factor upto the order of $10^4$, while the training dataset has Q-factor upto the order of $10^3$.

Figures

Figures reproduced from arXiv: 2507.18127 by the authors.

Figure 1
Figure 1. Metal-insulator-metal (MIM) and hybrid dielectric metasurface unit cells demonstrate distinct absorption [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. RGB encoding of metasurface unit cells, where geometry is embedded and material properties, thickness, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Schematic Training Architecture of a Hybrid Quantum Generative Adversarial Network (QGAN) with a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Testing and Validation Process of a Hybrid Quantum GAN: From Latent Space Sampling to the Generation of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the β-VAE and IWAE used in the paper. This study presents a comparative analysis of the outcomes derived from two pretrained variational autoencoders: the β-VAE and the importance weighted autoencoders (IWAE) [34, 35]. β-VAE and IWAE represent unsupervi…
Figure 6
Figure 6. Figure 6: Quantum neural network employed for the LaSt-QGAN model. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Absorption metasurface for both a, b) Lorentzian-shaped and c,d) Fano-shaped spectra from the validation [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Substitution of LaSt-QGAN-predicted materials with similar alternatives for absorption metasurfaces. a, [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Absorption spectra of metasurfaces exhibiting high-Q Fano resonances at target wavelengths. The inset [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 36 canonical work pages

  1. [14]

    Knight, and Aaswath P

    Christopher Yeung, Ryan Tsai, Benjamin Pham, Brian King, Yusaku Kawagoe, David Ho, Julia Liang, Mark W. Knight, and Aaswath P. Raman. Global inverse design across multiple photonic structure classes using generative deep learning. Advanced Optical Materials, 9(20):2100548, 2021

    work page 2021

  2. [1]

    Flat optics with designer metasurfaces

    Nanfang Yu and Federico Capasso. Flat optics with designer metasurfaces. Nature materials, 13(2):139–150, 2014

    work page 2014

  3. [2]

    Planar photonics with metasurfaces

    Alexander V Kildishev, Alexandra Boltasseva, and Vladimir M Shalaev. Planar photonics with metasurfaces. Science, 339(6125):1232009, 2013

    2013

  4. [3]

    Recent advances in planar optics: from plasmonic to dielectric metasurfaces

    Patrice Genevet, Federico Capasso, Francesco Aieta, Mohammadreza Khorasaninejad, and Robert Devlin. Recent advances in planar optics: from plasmonic to dielectric metasurfaces. Optica, 4(1):139–152, 2017

    work page 2017

  5. [4]

    Computational scaling in inverse photonic design through factorization caching

    Ahmet Onur Dasdemir, Victor Minden, and Emir Salih Magden. Computational scaling in inverse photonic design through factorization caching. Applied Physics Letters, 123(22):221106, 2023

    work page 2023

  6. [5]

    Inverse design in nanophotonics

    Sean Molesky, Zeyu Lin, Alexander Y Piggott, Weng Cho Chew Jin, Jelena Vuckovi´c, and Alejandro W Rodriguez. Inverse design in nanophotonics. Nature Photonics, 12:659–670, 2018

    work page 2018

  7. [6]

    Global optimization of dielectric metasurfaces using a physics-driven neural network

    Jicheng Jiang and Jonathan A Fan. Global optimization of dielectric metasurfaces using a physics-driven neural network. Nano letters, 19(8):5366–5372, 2019

    work page 2019

  8. [7]

    Training deep neural networks for the inverse design of nanophotonic structures

    Deyu Liu, Yichen Tan, Ehsan Khoram, and Zongfu Yu. Training deep neural networks for the inverse design of nanophotonic structures. ACS Photonics, 5(4):1365–1369, 2018

    work page 2018

Show all 45 references
  1. [8]

    Deep learning enabled inverse design in nanophotonics

    Sunae So, Trevon Badloe, Jaebum Noh, Jorge Bravo-Abad, and Junsuk Rho. Deep learning enabled inverse design in nanophotonics. Nanophotonics, 9(5):1041–1057, 2020. 11 arXiv Template A PREPRINT

    2020

  2. [9]

    Inverse molecular design using machine learning: Genera- tive models for matter engineering

    Benjamin Sanchez-Lengeling and Alán Aspuru-Guzik. Inverse molecular design using machine learning: Genera- tive models for matter engineering. Science, 361(6400):360–365, 2018

    2018

  3. [10]

    Kun-Hua Tu, Hejin Huang, Sangho Lee, Wonmoo Lee, Zehao Sun, Alfredo Alexander-Katz, and Caroline A. Ross. Machine learning predictions of block copolymer self-assembly. Advanced Materials, 32(52):2005713, 2020

    2020

  4. [11]

    Deep-learning-enabled fast optical identification and characterization of 2d materials.Advanced Materials, 32(29), 2020

    Bingnan Han et al. Deep-learning-enabled fast optical identification and characterization of 2d materials.Advanced Materials, 32(29), 2020

    2020

  5. [12]

    Simultaneous inverse design of materials and structures via deep learning: Demonstration of dipole resonance engineering using core–shell nanoparticles

    Sunae So, Jungho Mun, and Junsuk Rho. Simultaneous inverse design of materials and structures via deep learning: Demonstration of dipole resonance engineering using core–shell nanoparticles. ACS Applied Materials & Interfaces, 11(27):24264–24268, 2019

    2019

  6. [13]

    Adversarial autoencoders,

    Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders,

  7. [15]

    Mode collapse in generative adversarial networks: An overview

    Youssef Kossale, Mohammed Airaj, and Aziz Darouichi. Mode collapse in generative adversarial networks: An overview. In 2022 8th International Conference on Optimization and Applications (ICOA) , pages 1–6, 2022

    2022

  8. [16]

    Challenges and opportunities in quantum machine learning

    Marco Cerezo, Guillaume Verdon, Hsin-Yuan Huang, et al. Challenges and opportunities in quantum machine learning. Nature Computational Science, 2:567–576, 2022

    2022

  9. [17]

    Unsupervised semantic-preserving adversarial hashing for image search

    Cheng Deng et al. Unsupervised semantic-preserving adversarial hashing for image search. IEEE Transactions on Image Processing, 28:4032–4044, 2019

    2019

  10. [18]

    Quantum Computing in the NISQ Era and Beyond

    John Preskill. Quantum Computing in the NISQ Era and Beyond. Quantum, 2:79, 2018

    2018

  11. [19]

    Quantum generative adversarial learning

    Seth Lloyd and Christian Weedbrook. Quantum generative adversarial learning. Physical Review Letters , 121(4):040502, 2018

    2018

  12. [20]

    Quantum generative adversarial networks

    Pierre-Luc Dallaire-Demers and Nathan Killoran. Quantum generative adversarial networks. Physical Review A, 98(1):012324, 2018

    2018

  13. [21]

    Experimental Quantum Generative Adversarial Networks for Image Generation

    He-Liang Huang, Yuxuan Du, Ming Gong, Youwei Zhao, Yulin Wu, Chaoyue Wang, Shaowei Li, Futian Liang, Jin Lin, Yu Xu, Rui Yang, Tongliang Liu, Min-Hsiu Hsieh, Hui Deng, Hao Rong, Cheng-Zhi Peng, Chao-Yang Lu, Yu-Ao Chen, Dacheng Tao, Xiaobo Zhu, and Jian-Wei Pan. Experimental Q...

    2021

  14. [22]

    West, Sarah M

    Shu Lok Tsang, Maxwell T. West, Sarah M. Erfani, and Muhammad Usman. Hybrid quantum–classical generative adversarial network for high-resolution image generation. IEEE Transactions on Quantum Engineering, 4:1–19, 2023

    2023

  15. [23]

    Stein, Betis Baheri, Daniel Chen, Ying Mao, Qiang Guan, Ang Li, Bo Fang, and Shuai Xu

    Samuel A. Stein, Betis Baheri, Daniel Chen, Ying Mao, Qiang Guan, Ang Li, Bo Fang, and Shuai Xu. Qugan: A quantum state fidelity based generative adversarial network. In 2021 IEEE International Conference on Quantum Computing and Engineering (QCE) , pages 71–81, 2021

    2021

  16. [24]

    Variational quantum generators: Generative adversarial quantum machine learning for continuous distributions

    Jonathan Romero and Alán Aspuru-Guzik. Variational quantum generators: Generative adversarial quantum machine learning for continuous distributions. Advanced Quantum Technologies, 4(1):2000003, 2021

    2021

  17. [25]

    Latent style-based quantum gan for high-quality image generation, 2024

    Su Yeon Chang, Supanut Thanasilp, Bertrand Le Saux, Sofia Vallecorsa, and Michele Grossi. Latent style-based quantum gan for high-quality image generation, 2024. Preprint at https://arxiv.org/abs/2406.02668

    2024 arXiv

  18. [26]

    DCGAN Tutorial – PyTorch Tutorials

    Nathan Inkawhich. DCGAN Tutorial – PyTorch Tutorials. https://pytorch.org/tutorials/beginner/ dcgan_faces_tutorial.html, 2022. Accessed: May 2022

    2022

  19. [27]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes, 2013. Preprint at https://doi.org/ 10.48550/arXiv.1312.6114

    2013 doi arXiv

  20. [28]

    X. Liu, T. Tyler, T. Starr, A. F. Starr, N. M. Jokerst, and W. J. Padilla. Taming the blackbody with infrared metamaterials as selective thermal emitters. Phys. Rev. Lett., 107(4):045901, 2011

    2011

  21. [29]

    Neutens, P

    P. Neutens, P. Van Dorpe, I. De Vlaminck, L. Lagae, and G. Borghs. Electrical detection of confined plasmonic resonances in metallic waveguides. Nat. Photonics, 3:283–286, 2009

    2009

  22. [30]

    S. Chen, Z. Chen, J. Liu, J. Cheng, Y . Zhou, L. Xiao, and K. Chen. Ultra-narrow band mid-infrared perfect absorber based on hybrid dielectric metasurface. Nanomaterials, 9(10):1350, 2019

    2019

  23. [31]

    Multiclass Metasurface Inverse Design

    Raman-Lab-UCLA. Multiclass Metasurface Inverse Design. https://github.com/Raman-Lab-UCLA/ Multiclass_Metasurface_InverseDesign/tree/main, 2024. Accessed: May 2022. 12 arXiv Template A PREPRINT

    2024

  24. [32]

    Fano resonances in photonics

    Mikhail Limonov, Mikhail Rybin, Alexander Poddubny, and Yuri Kivshar. Fano resonances in photonics. Nature Photonics, 11:543–554, 2017

    2017

  25. [33]

    COMSOL Multiphysics® v

    COMSOL AB, Stockholm, Sweden. COMSOL Multiphysics® v. 6.2, 2023. Accessed: 2024-05-29

    2023

  26. [34]

    beta-V AE: Learning Basic Visual Concepts with a Constrained Variational Framework

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-V AE: Learning Basic Visual Concepts with a Constrained Variational Framework. In International Conference on Learning Representations , p...

    2017

  27. [35]

    Importance Weighted Autoencoders, 2015

    Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov. Importance Weighted Autoencoders, 2015. Preprint at https://arxiv.org/abs/1509.00519

    2015 arXiv

  28. [36]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

    2019

  29. [37]

    Pennylane: Automatic differentiation of hybrid quantum-classical computations, 2018

    Ville Bergholm et al. Pennylane: Automatic differentiation of hybrid quantum-classical computations, 2018. Preprint at https://arxiv.org/abs/1811.04968

    2018 arXiv

  30. [38]

    Grabowska, and Stefano Carrazza

    Carlos Bravo-Prieto, Jacopo Baglio, Marco Cè, Anthony Francis, Dorota M. Grabowska, and Stefano Carrazza. Style-based quantum generative adversarial networks for monte carlo events. Quantum, 6:777, 2022

    2022

  31. [39]

    Variational quantum algorithms

    Marco Cerezo et al. Variational quantum algorithms. Nature Reviews Physics, 3(9):625–644, 2021

    2021

  32. [40]

    Love, Alán Aspuru- Guzik, and Jeremy L

    Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter J. Love, Alán Aspuru- Guzik, and Jeremy L. O’Brien. A variational eigenvalue solver on a photonic quantum processor. Nature Communications, 5:4213, 2014

    2014

  33. [41]

    Classical and quantum algorithms for orthogonal neural networks, 2021

    Iordanis Kerenidis, Jonas Landman, and Natansh Mathur. Classical and quantum algorithms for orthogonal neural networks, 2021. Preprint at https://arxiv.org/abs/2106.07198

    2021 arXiv

  34. [42]

    Reluplex made more practical: Leaky ReLU

    Jin Xu, Zishan Li, Bowen Du, Miaomiao Zhang, and Jing Liu. Reluplex made more practical: Leaky ReLU. In 2020 IEEE Symposium on Computers and Communications (ISCC) , pages 1–7, 2020

    2020

  35. [43]

    Approximation by superpositions of a sigmoidal function

    George Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals and Systems, 2:303–314, 1989

    1989

  36. [44]

    M. N. Polyanskiy. Refractive index database. https://refractiveindex.info, 2024. (accessed: Aug 2024). 13

    2024

  37. [2015]

    Preprint at https://arxiv.org/abs/1511.05644

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.