Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Multifunctional Metasurface Design with a Generative Adversarial Network

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A conditional Wasserstein GAN with a prediction-network filter can generate free-form meta-atoms meeting amplitude and phase targets, and assembled metasurfaces verify them in simulation.

desk verdict A real advance for GAN-based metasurface design, with device-level full-wave validation, but the quantitative pass-rate claims rest on an unquantified, out-of-distribution surrogate PNN screening step. read the letter →

arxiv 1908.04851 v2 pith:IUPRDE6S submitted 2019-08-13 physics.optics cs.LG

Sensong An , Bowen Zheng , Hong Tang , Mikhail Y. Shalaginov , Li Zhou , Hang Li , Tian Gu , Juejun Hu
show 2 more authors
Clayton Fowler Hualiang Zhang
This is my paper · ORCID
classification physics.opticscs.LG
keywords metasurfaceinversedesigndeeplearninggenerativeadversarialnetworkWassersteinGANall-dielectricmeta-atomsmultifunctionalpolarizationmultiplexing
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

One of the hardest parts of building multifunctional metasurfaces is the inverse design step: finding individual meta-atom shapes whose transmission amplitude and phase are exactly what the device needs, when several independent requirements must be met at once. This paper claims that the problem can be turned into a generative one. A conditional Wasserstein GAN, trained once on about 29,000 full-wave-simulated free-form meta-atoms, draws new meta-atom images conditioned on the desired complex transmission coefficient, with a pretrained prediction network screening out designs that miss the target; the same architecture handles dual-polarization goals by feeding in four targets instead of two. The authors verify the claim by assembling GAN-generated meta-atoms into a bifocal metalens, a polarization-multiplexed beam deflector, a polarization-multiplexed metalens, and a polarization-independent metalens, all of which show the intended behavior in full-wave simulation.

What carries the argument

The load-bearing mechanism is the conditional Wasserstein GAN (cWGAN) itself: a generator that turns a condition vector plus noise into a 64 × 64 binary meta-atom image, and a discriminator that enforces fidelity by maximizing the Wasserstein distance between real and generated samples. Two details make it work for optics. First, the condition vector stores the real and imaginary parts of the desired complex transmission coefficient rather than raw amplitude and phase, which avoids the phase-jump discontinuities that plague direct phase conditioning; for multifunctional designs the vector is simply concatenated across polarization channels. Second, a customized geometry-aware gradient penalty replaces numeric interpolation with spatial splicing of real and fake binary images, keeping the interpolated samples physical while enforcing the Lipschitz constraint that stabilizes WGAN training. In deployment a pretrained prediction neural network (PNN) estimates each generated image's transmission spectrum and discards out-of-tolerance candidates before any full-wave simulation is run.

What would settle it

Take a fresh batch of GAN-generated meta-atoms that the prediction network accepts, choose those whose binary images contain features smaller than the 0.1 μm resolution of the training dataset, and compute their amplitude and phase by full-wave simulation; if the error rate is substantially worse than the paper's reported one-in-600, the filter's surrogate predictions—not the GAN's generative accuracy—are carrying the result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that meta-atom inverse design can be reformulated as conditional image generation: after training, the generator maps a vector of desired transmission coefficients (real and imaginary parts, for one or two orthogonal polarizations) plus random noise directly to a 64 × 64 binary meta-atom pattern, in a single forward pass and without iterative search. In single-target tests the authors report that 599 of 600 generated meta-atoms fell within ±0.1 in amplitude and ±10° in phase when checked by full-wave simulation, and the dual-polarization network produced meta-atoms meeting four simultaneous targets. Assembling those meta-atoms into four distinct device classes showed that this accuracy propagates to the device level: full-wave simulations of the lenses and deflector reproduce the targeted focal lengths, focal shifts, and deflection angles. The authors therefore claim to demonstrate the first free-form all-dielectric meta-atom design network, the first free-form multifunctional metasurface design network, and the first metasurface lens designed by a GAN.

Load-bearing premise

The load-bearing assumption is that the pretrained prediction network correctly estimates the transmission of any shape the GAN produces, including free-form patterns outside the 29,000 needle-drop training examples; if the surrogate is overconfident on novel geometry, the reported pass rates and the claimed time savings would be optimistic.

Editorial extensions

If this is right

  • After one training run, a target amplitude-and-phase mask can be populated with hundreds of qualified free-form meta-atoms in seconds, removing per-device iterative optimization from the design loop.
  • Because multifunctionality is expressed by enlarging the condition vector, the same trained procedure extends to polarization-multiplexed devices and, in principle, to multi-wavelength, angular, or tunable-material targets without architectural changes.
  • The generator produces shapes that were not in the 28 × 28 training data, such as inclined edges and rounded corners, which gives designers a pool of structurally distinct but electromagnetically equivalent options for fabrication-tolerance selection.
  • The generated designs can seed local optimizers, offering a route past the initial-guess sensitivity and local-minima problems of evolutionary or gradient-based search.
  • Clusters of generated designs with identical electromagnetic responses can be mined for shared geometry, turning the network into a tool for discovering the physical origin of a given response.

Reading between the lines

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

  • The paper does not report how many valid designs the prediction network falsely discards, so the end-to-end efficiency of the GAN alone relative to evolutionary search is unresolved; a fair comparison would count every PNN call, not just the accepted designs.
  • If the surrogate's accuracy carries over to out-of-distribution shapes, the same condition-vector construction should extend to wavelength-multiplexed and angle-multiplexed metasurfaces; the paper lists these as possible extensions but does not test them.
  • The generator's 64 × 64 output can draw features below the 0.1 μm grid on which the needle-drop training patterns were defined, so practical adoption would require rasterizing or rounding those sub-resolution details and re-checking them, a step the paper does not address.
  • If the approach scales to larger apertures, it opens the possibility of real-time re-targetable flat optics, since a new phase mask can be populated immediately after training without simulation or optimization.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a conditional Wasserstein Generative Adversarial Network (WGAN) with a gradient penalty for inverse design of free-form all-dielectric meta-atoms, conditioned on complex transmission coefficients (amplitude and phase). The generator output is screened by a pre-trained prediction neural network (PNN) before designs are accepted. The authors demonstrate the method by generating meta-atoms for single-target and dual-polarization targets, and by assembling a bifocal metalens, a polarization-multiplexed beam deflector, a polarization-multiplexed metalens, and a polarization-independent metalens. Final devices are validated with independent full-wave simulations.

Significance. If the central claims hold, this would be a noteworthy advance: it extends GAN-based inverse design to free-form all-dielectric meta-atoms with simultaneous amplitude and phase constraints, and it demonstrates multifunctional devices with full-wave verification. The paper is also honest about the training-data generation and provides detailed network architectures. The main weakness is that the quantitative success rates are reported after screening by the authors' own PNN, and the PNN's error is not quantified, while the GAN is credited with generating structures outside the PNN's training distribution. These issues do not invalidate the full-wave-verified device demonstrations, but they weaken the headline claim that the GAN itself is generating on-demand designs and the aggregate accuracy statistics.

major comments (3)
  1. [Meta-atom design network / Fig. 2] The reported success statistics (e.g., 'one unqualified design among 600') are computed on designs that have already been screened by the PNN, not on the raw outputs of the GAN. The manuscript states that the PNN 'characterizes these output images and eliminates the unqualified meta-atom designs' (p. 7), and it does not report the raw generation success rate or quantify the PNN's prediction error. This matters because the PNN was trained on needle-drop patterns composed of rectangles with 0.1 μm resolution and 4-fold symmetry (SI Section 2), whereas the GAN is credited with generating inclined edges and round corners 'not included in the training data' (p. 17). The PNN is therefore being used to score exactly the structures on which a surrogate is least reliable. To support the central claim that the GAN itself generates on-demand designs, please report raw GAN pass rates before PNN filtering, provide PNN validation accuracy (ideally on held-out and out-of-distribution geometries), and full-wave-verify a random sample of PNN-filtered and PNN-rejected designs.
  2. [Discussion and conclusion (p. 16)] The claim that the approach is 'much more effective' than trial-and-error or global optimization is not backed by any baseline comparison. Without a quantitative comparison against, e.g., random needle-drop sampling, an evolutionary algorithm, or a direct optimization method evaluated with the same full-wave verification, the advantage over existing approaches is asserted rather than demonstrated. Please add at least one baseline that uses the same data and the same qualification thresholds, reporting success rate per unit time and device-level performance.
  3. [Figs. 3, 4, 5 and SI Section 5] The device-level validation is qualitative. The text states 'excellent agreement' between targets and simulated masks and shows focal spots, but it does not provide quantitative metrics such as focal-spot efficiency, Strehl ratio, diffraction efficiency for the deflector (Fig. 4g-h), or the ratio of focal intensity to background. Adding these numbers would directly substantiate the paper's accuracy claims and would also make the comparison to future work meaningful.
minor comments (6)
  1. [p. 5 and SI Section 1] The main text says the generator has 'seven consecutive transposed convolutional layers,' while the SI says 'eight consecutive transposed convolution layers'; please reconcile.
  2. [p. 8 and SI Table S1] The main text says training stabilized after 1,500 epochs, but the SI reports 3,000 iterations (72 h); please clarify which number corresponds to the models used for Figs. 2–5.
  3. [Fig. 2 and p. 8] The phrase 'we employed the well trained GAN to consecutively generate 100 qualified designs' is ambiguous; state whether the 100 are raw samples or post-PNN-selected samples.
  4. [Eq. (4)] There is a typesetting error ('𝑒𝑥𝑒𝑒𝑗2𝜋𝜋/𝜆𝜆 𝑑𝑑1'); the exponential should be written as exp(j2πd1/λ) or similar.
  5. [p. 7 / SI Section 1] The generator uses a tanh output layer, but binarization of output patterns is not described; please explain how continuous generator outputs are converted to binary meta-atom patterns before PNN evaluation and full-wave simulation.
  6. [Reference [49]] The PNN is cited as a preprint and is not described in enough detail for independent reproduction; please provide a public repository or complete architectural and training details beyond what is in the SI.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: full-wave simulations independently validate the GAN-generated devices; the self-cited PNN is a screening surrogate, not the final evidence.

full rationale

The paper's load-bearing demonstrations are validated by independent full-wave simulations, not by the network's own outputs or by the self-cited prediction network. The GAN generates candidate meta-atoms, the PNN (ref. [49]) screens them, and then the electromagnetic responses of the surviving designs are computed with full-wave tools: 'The electromagnetic responses of the generated patterns were computed using full-wave simulation tools and are labeled with blue dots' (Fig. 2). The final devices are likewise verified by CST full-wave simulations, e.g., the bifocal metalens field distribution in Fig. 3f and the polarization-multiplexed metalens fields in Fig. 5e-f. The PNN is used to speed up rejection of unpromising candidates, but the reported 'one unqualified design among 600' and the device-level agreement are based on full-wave data, so the success rates are not determined by construction from the PNN. The self-citation to [49] is real but not load-bearing in a circular way: even if the PNN were imperfect, the full-wave verification of final devices stands independently. The absence of a raw GAN success rate is a reporting/evidence limitation concerning generalization, not a case where a prediction reduces to its fitted input. No equation is shown to equal its own input, no fitted parameter is renamed as a prediction, and no uniqueness claim is imported from prior work to force the design choice.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No physical constants are fitted. The reported success depends on hand-selected tolerances, training-data generation rules, the accuracy of the self-cited PNN surrogate, and the empirical stability of a custom gradient-penalty interpolation that lacks theoretical justification.

free parameters (3)
  • Qualification thresholds for meta-atom pass/fail = amplitude error +/- 0.1, phase error +/- 10 degrees
    Hand-selected tolerances define every reported pass rate and the qualification statistics in Figs. 2 and 4; looser or tighter thresholds would change the claimed accuracy.
  • Needle-drop training data generation parameters = 3 to 7 randomly placed needles, 0.4 micron minimum spacing, 0.1 micron feature resolution
    These choices define the training distribution and therefore constrain what the GAN and PNN can produce; the paper does not test sensitivity to them.
  • WGAN-GP penalty coefficient lambda = 10
    Standard hand-chosen hyperparameter for the gradient-penalty loss; it affects training stability but is not a physical parameter fitted to the devices.
assumptions (4)
  • domain assumption Full-wave CST simulations give the ground-truth electromagnetic response for all meta-atoms.
    No experimental measurement appears anywhere in the paper; the entire pipeline is validated against CST, so any systematic CST error enters every claim.
  • domain assumption The PNN surrogate (ref. [49], same group) predicts the complex transmission coefficients of arbitrary free-form meta-atoms accurately enough to filter designs.
    The paper does not quantify PNN error in the main text, yet uses it to discard unqualified designs before full-wave verification.
  • ad hoc to paper The custom geometry interpolation in SI Section 3 preserves the 1-Lipschitz constraint and training stability of WGAN-GP.
    Replacing straight-line interpolation with random cut-and-paste of binarized patterns is not shown to satisfy the gradient-penalty theory; training stability is demonstrated empirically only.
  • domain assumption The 29,000 retained training patterns are representative of the target free-form design space for the demonstrated multifunctional devices.
    No coverage analysis of the phase-amplitude-condition space is provided; the network can only compose designs from this training manifold.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multifunctional Metasurface Design with a Generative Adversarial Network." pith.science (2026). https://pith.science/paper/IUPRDE6S

@misc{pith2026190804851,
  author       = {Pith},
  title        = {Pith review of: Multifunctional Metasurface Design with a Generative Adversarial Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUPRDE6S}},
  note         = {Machine review of arXiv:1908.04851}
}
read the original abstract

Metasurfaces have enabled precise electromagnetic wave manipulation with strong potential to obtain unprecedented functionalities and multifunctional behavior in flat optical devices. These advantages in precision and functionality come at the cost of tremendous difficulty in finding individual meta-atom structures based on specific requirements (commonly formulated in terms of electromagnetic responses), which makes the design of multifunctional metasurfaces a key challenge in this field. In this paper, we present a Generative Adversarial Networks (GAN) that can tackle this problem and generate meta-atom/metasurface designs to meet multifunctional design goals. Unlike conventional trial-and-error or iterative optimization design methods, this new methodology produces on-demand free-form structures involving only a single design iteration. More importantly, the network structure and the robust training process are independent of the complexity of design objectives, making this approach ideal for multifunctional device design. Additionally, the ability of the network to generate distinct classes of structures with similar electromagnetic responses but different physical features could provide added latitude to accommodate other considerations such as fabrication constraints and tolerances. We demonstrate the network's ability to produce a variety of multifunctional metasurface designs by presenting a bifocal metalens, a polarization-multiplexed beam deflector, a polarization-multiplexed metalens and a polarization-independent metalens.

Figures

Figures reproduced from arXiv: 1908.04851 by the authors.

Figure 1
Figure 1. Network architecture of the generative meta-atom design network. (a) Schematic diagram of the proposed network. The discriminator network measures the Wasserstein distance between real and fake samples and aims to maximize the distance between them. The generator network attempts to confuse the discriminator by transforming target conditions combined with noise prior to producing fake samples that resemble real ones… view at source ↗
Figure 2
Figure 2. Meta-atom designs generated using a fully-trained conditional WGAN model. 100 meta-atom designs were produced for each combined condition of amplitude and phase: (a) 0.5 + 180°, (b) 0.9 + 90°, (c) 0.7 + 0°, (d) 0.6 + 45°, (e) 0.3 + 200° and (f) 0.8 + 180°, respectively. In the polar charts the radial and angular coordinates correspond to the transmission amplitude and phase. Blue dots represent EM responses of gener… view at source ↗
Figure 3
Figure 3. A bifocal lens designed with the meta-atom generative network. (a) Target amplitude mask and (b) Target phase mask of the designed lateral bifocal metalens at 50 THz with f1 = f2 = 60 μm. (c) Theoretical Ex field distributions in the X-Z plane calculated with MATLAB. (d) Metasurface pattern designed by the meta-atom generative network. Since the target E-field is symmetrical along x and y axes, only one quadrant of … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 55 canonical work pages

  1. [1]

    H. C. Hulst, H. C. van de Hulst, Light scattering by small particles, Courier Corporation, 1981

    work page 1981

  2. [2]

    J. D. Jackson, Classical electrodynamics, John Wiley & Sons, 2007

    work page 2007

  3. [3]

    Q. Wang, X. Zhang, Y. Xu, J. Gu, Y. Li, Z. Tian, R. Singh, S. Zhang, J. Han, W. Zhang, Scientific reports 2016, 6, 32867

    work page 2016

  4. [4]

    Arbabi, Y

    A. Arbabi, Y. Horie, M. Bagheri, A. Faraon, Nature nanotechnology 2015, 10, 937

    work page 2015

  5. [5]

    N. Yu, F. Capasso, Nature materials 2014, 13, 139

    work page 2014

  6. [6]

    Jahani, Z

    S. Jahani, Z. Jacob, Nature nanotechnology 2016, 11, 23

    work page 2016

  7. [7]

    Khorasaninejad, Z

    M. Khorasaninejad, Z. Shi, A. Y. Zhu, W.-T. Chen, V. Sanjeev, A. Zaidi, F. Capasso, Nano letters 2017, 17, 1819

    work page 2017

  8. [8]

    X. Wang, J. Ding, B. Zheng, S. An, G. Zhai, H. Zhang, Scientific reports 2018, 8, 1876

    work page 2018

Show all 57 references
  1. [9]

    J. Ding, S. An, B. Zheng, H. Zhang, Advanced Optical Materials 2017, 5, 1700079

    2017

  2. [10]

    T. Wang, R. Xie, S. Zhu, J. Gao, M. Xin, S. An, B. Zheng, H. Li, Y. Lin, H. Zhang, IEEE Access 2019, 7, 58131

    2019

  3. [11]

    L. Liu, X. Zhang, M. Kenney, X. Su, N. Xu, C. Ouyang, Y. Shi, J. Han, W. Zhang, S. Zhang, Advanced materials 2014, 26, 5031

    2014

  4. [12]

    D. Sell, J. Yang, S. Doshay, J. A. Fan, Advanced Optical Materials 2017, 5, 1700645

    2017

  5. [13]

    Cheng, X

    H. Cheng, X. Wei, P. Yu, Z. Li, Z. Liu, J. Li, S. Chen, J. Tian, Applied Physics Letters 2017, 110, 171903

    2017

  6. [14]

    Callewaert, V

    F. Callewaert, V. Velev, P. Kumar, A. Sahakian, K. Aydin, Scientific reports 2018, 8, 1358

    2018

  7. [15]

    C. H. Chu, M. L. Tseng, J. Chen, P. C. Wu, Y. H. Chen, H. C. Wang, T. Y. Chen, W. T. Hsieh, H. J. Wu, G. Sun, Laser & Photonics Reviews 2016, 10, 986

    2016

  8. [16]

    M. Y. Shalaginov, S. An, Y. Zhang, F. Yang, P. Su, V. Liberman, J. B. Chou, C. M. Roberts, M. Kang, C. Rios, arXiv preprint arXiv:1911.12970 2019

    1911 arXiv

  9. [17]

    Zhang, J

    Y. Zhang, J. B. Chou, J. Li, H. Li, Q. Du, A. Yadav, S. Zhou, M. Y. Shalaginov, Z. Fang, H. Zhong, Nature communications 2019, 10, 1

    2019

  10. [18]

    R. C. Devlin, M. Khorasaninejad, W. T. Chen, J. Oh, F. Capasso, Proceedings of the National Academy of Sciences 2016, 113, 10473

    2016

  11. [19]

    Arbabi, A

    E. Arbabi, A. Arbabi, S. M. Kamali, Y. Horie, A. Faraon, Optica 2016, 3, 628

    2016

  12. [20]

    D. Z. Zhu, E. B. Whiting, S. D. Campbell, D. B. Burckel, D. H. Werner, ACS Photonics 2019, 6, 2741

    2019

  13. [21]

    High-Performance Metasurfaces Synthesized via Multi-Objective Optimization

    S. D. Campbell, E. B. Whiting, D. H. Werner, P. L. Werner, "High-Performance Metasurfaces Synthesized via Multi-Objective Optimization", presented at 2019 International Applied Computational Electromagnetics Society Symposium (ACES), 2019

    2019

  14. [22]

    D. Liu, Y. Tan, E. Khoram, Z. Yu, ACS Photonics 2018, 5, 1365

    2018

  15. [23]

    Z. Liu, D. Zhu, S. P. Rodrigues, K.-T. Lee, W. Cai, Nano letters 2018, 18, 6570

    2018

  16. [24]

    W. Ma, F. Cheng, Y. Liu, ACS nano 2018, 12, 6326

    2018

  17. [25]

    Malkiel, M

    I. Malkiel, M. Mrejen, A. Nagler, U. Arieli, L. Wolf, H. Suchowski, Light: Science & Applications 2018, 7, 60

    2018

  18. [26]

    Peurifoy, Y

    J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B. G. DeLacy, J. D. Joannopoulos, M. Tegmark, M. Soljačić, Science advances 2018, 4, eaar4206

    2018

  19. [27]

    S. An, C. Fowler, B. Zheng, M. Y. Shalaginov, H. Tang, H. Li, L. Zhou, J. Ding, A. M. Agarwal, C. Rivero-Baleine, ACS Photonics 2019, 6, 3196

    2019

  20. [28]

    S. So, J. Mun, J. Rho, ACS applied materials & interfaces 2019

    2019

  21. [29]

    Deep neural network (DNN) surrogate models for the accelerated design of optical devices and systems

    R. S. Hegde, "Deep neural network (DNN) surrogate models for the accelerated design of optical devices and systems", presented at Novel Optical Systems, Methods, and Applications XXII, 2019. 19

    2019

  22. [30]

    Z. Liu, Z. Zhu, W. Cai, Optics Express 2020, 28, 4825

    2020

  23. [31]

    Inverse Design of Photonic Crystal Nanobeam Cavity Structure via Deep Neural Network

    J. Hao, L. Zheng, D. Yang, Y. Guo, "Inverse Design of Photonic Crystal Nanobeam Cavity Structure via Deep Neural Network", presented at Asia Communications and Photonics Conference, 2019

    2019

  24. [32]

    Lin, C.-C

    K.-F. Lin, C.-C. Hsieh, S.-C. Hsin, W.-F. Hsieh, Applied Optics 2019, 58, 8914

    2019

  25. [33]

    X. Han, Z. Fan, C. Li, Z. Liu, L. J. Guo, arXiv preprint arXiv:1912.03696 2019

    1912 arXiv

  26. [34]

    Generative adversarial nets

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, "Generative adversarial nets", presented at Advances in neural information processing systems, 2014

    2014

  27. [35]

    Mirza, S

    M. Mirza, S. Osindero, arXiv preprint arXiv:1411.1784 2014

    2014 arXiv

  28. [36]

    Jiang, J

    J. Jiang, J. A. Fan, arXiv preprint arXiv:1906.07843 2019

    1906 arXiv

  29. [37]

    Jiang, J

    J. Jiang, J. A. Fan, Nano letters 2019, 19, 5366

    2019

  30. [38]

    Jiang, D

    J. Jiang, D. Sell, S. Hoyer, J. Hickey, J. Yang, J. A. Fan, ACS nano 2019, 13, 8872

    2019

  31. [39]

    W. Ma, F. Cheng, Y. Xu, Q. Wen, Y. Liu, Advanced Materials 2019, 31, 1901111

    2019

  32. [40]

    S. So, J. Rho, Nanophotonics 2019, 8, 1255

    2019

  33. [41]

    Joint multi-layer GAN-based design of tensorial RF metasurfaces

    J. A. Hodge, K. V. Mishra, A. I. Zaghloul, "Joint multi-layer GAN-based design of tensorial RF metasurfaces", presented at 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP), 2019

    2019

  34. [42]

    Multidiscriminator distributed generative model for multi-layer RF metasurface discovery

    J. A. Hodge, K. V. Mishra, A. I. Zaghloul, "Multidiscriminator distributed generative model for multi-layer RF metasurface discovery", presented at IEEE Global Conference on Signal and Information Processing, 2019

    2019

  35. [43]

    Sajedian, J

    I. Sajedian, J. Kim, J. Rho, Microsystems & nanoengineering 2019, 5, 1

    2019

  36. [44]

    Z. Liu, D. Zhu, K.-T. Lee, A. S. Kim, L. Raju, W. Cai, arXiv preprint arXiv:1907.03366 2019

    1907 arXiv

  37. [45]

    Wasserstein generative adversarial networks

    M. Arjovsky, S. Chintala, L. Bottou, "Wasserstein generative adversarial networks", presented at International conference on machine learning, 2017

    2017

  38. [46]

    Improved training of wasserstein gans

    I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, A. C. Courville, "Improved training of wasserstein gans", presented at Advances in neural information processing systems, 2017

    2017

  39. [47]

    Delving deep into rectifiers: Surpassing human- level performance on imagenet classification

    K. He, X. Zhang, S. Ren, J. Sun, "Delving deep into rectifiers: Surpassing human- level performance on imagenet classification", presented at Proceedings of the IEEE international conference on computer vision, 2015

    2015

  40. [48]

    Ioffe, C

    S. Ioffe, C. Szegedy, arXiv preprint arXiv:1502.03167 2015

    2015 arXiv

  41. [49]

    S. An, B. Zheng, M. Y. Shalaginov, H. Tang, H. Li, L. Zhou, J. Ding, A. M. Agarwal, C. Rivero-Baleine, M. Kang, arXiv preprint arXiv:2001.00121 2020

    2001 arXiv

  42. [50]

    J. Ding, N. Xu, H. Ren, Y. Lin, W. Zhang, H. Zhang, Scientific reports 2016, 6, 1

    2016

  43. [51]

    Kiarashinejad, S

    Y. Kiarashinejad, S. Abdollahramezani, M. Zandehshahvar, O. Hemmatyar, A. Adibi, arXiv preprint arXiv:1905.06889 2019

    1905 arXiv

  44. [52]

    Kiarashinejad, M

    Y. Kiarashinejad, M. Zandehshahvar, S. Abdollahramezani, O. Hemmatyar, R. Pourabolghasem, A. Adibi, arXiv preprint arXiv:1909.07330 2019. 20 Supporting Information Title: Multifunctional Metasurface Design with a Generative Adversarial Network Sensong An 1, Bowen Zheng1, Hong ...

    1909 arXiv

  45. [53]

    S1 illustrates the network architecture of the proposed neural network, with the data flow details included

    Detailed network architecture Fig. S1 illustrates the network architecture of the proposed neural network, with the data flow details included. Input conditions, such as frequency-dependent amplitude and phase responses, polarization- dependence, and/or material states ( e.g. ...

  46. [54]

    needle d rop

    Training data collection Without loss of generality, the all -dielectric meta-atom consists of a 1 μm thick dielectric component (preferably with a high refractive index, n1. In this case n1 = 5) sitting on a dielectric substrate (preferably with a low refractive index, n2. In...

  47. [55]

    1” and black represents “0

    Customized gradient-penalty method The Wasserstein distance is only accurate when the discriminator is a 1-Lipschitz function.[45] To enforce this constraint, the original WGAN applied a simple, but rough, value clipping to restrict the maximum weight value in each layer of th...

  48. [56]

    Hyperparameters used in the training of GANs and PNNs

    Hyperparameters and training curves Table S1. Hyperparameters used in the training of GANs and PNNs. Hyperparameters Meta-atom design network Dual-polarization meta-atom design network PNN for real part PNN for imaginary part Training set size 29,000 29,000 69,000 69,000 Optim...

  49. [57]

    A polarization -independent focusing lens designed with the dual -polarization meta -atom generative network

    Polarization-independent metalens design Figure S7 . A polarization -independent focusing lens designed with the dual -polarization meta -atom generative network. (a) Metasurface pattern designed with the dual -polarization meta-atom generative network. Part of the metasurface...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.