Inverse Design using Physics-Informed Quantum GANs for Tailored Absorption in Dielectric Metasurfaces

Jayasri Dontabhaktuni; Sreeraj Rajan Warrier

REVIEW 4 major objections 3 minor 23 references

Inverse Design using Physics-Informed Quantum GANs for Tailored Absorption in Dielectric Metasurfaces

T0 review · 4 major / 3 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 By adding a Fano-resonance physics loss to a hybrid quantum-classical GAN, this paper claims inverse metasurface design succeeds with roughly 100 training samples, order-of-magnitude lower error, faster training, and generated quality…

desk verdict A plausible new combination of QGAN and physics-informed loss, but the headline comparisons are uncontrolled and the high-Q claim is under-validated; the idea deserves a proper test, not this evidence. read the letter →

arxiv 2507.18132 v1 pith:4MHZGB4X submitted 2025-07-24 physics.optics cond-mat.soft

Sreeraj Rajan Warrier , Jayasri Dontabhaktuni This is my paper

classification physics.opticscond-mat.soft

keywords metasurfacesnarrow-bandabsorptioninversedesignquantumgenerativeadversarialnetworkphysics-informedneuralnetworksFanoresonancequalityfactorvariationalcircuit

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

The paper sets out to solve an inverse design problem: given a target narrow-band Fano absorption profile (central frequency, linewidth, peak absorption, and asymmetry), produce a dielectric metasurface unit cell whose real spectrum matches it. The proposed solution is a generative adversarial network whose loss function is augmented with the Fano absorption formula, so the generator is pushed toward physically consistent geometries. The authors compare the physics-informed classical GAN with a version whose generator is a variational quantum circuit, and report that the quantum version converges faster, uses only about 100 training samples versus 20,000 for the classical GAN baseline, and lowers the mean squared error from about $10^{-3}$ to $10^{-6}$. They also report that adding a Q-factor floor of $10^{5}$ to the loss yields asymmetric unit cells with simulated quality factors up to 1.56×$10^{5}$, even though the training set only contained resonances with Q near $10^{3}$. If these numbers hold, the paper would demonstrate that quantum-enhanced, physics-informed generative models can explore high-dimensional photonic design spaces under severe data scarcity.

What carries the argument

The load-bearing machinery is the Fano absorption line-shape $A(\nu)=A_0(q+\delta)^2/(1+\delta^2)$ with $\delta=2(\nu-\nu_0)/\Gamma$, converted into a physics-informed loss $L_{\text{physics}}=\lambda_A L_A+\lambda_{\nu_0} L_{\nu_0}+\lambda_Q L_Q$, where $L_A$ penalizes absorption below target, $L_{\nu_0}$ penalizes resonance-frequency deviation, and $L_Q$ penalizes quality factors below a threshold. The discriminator is simultaneously an adversary and a physics surrogate: it predicts the four Fano parameters from the generated image, so the physics loss is computed from predicted spectra rather than from a full electromagnetic simulation. For the quantum version, the generator is a variational quantum circuit with $RY$ rotations and cyclic $CNOT$ entanglements, whose Pauli-$Z$ expectation values pass through a pretrained importance-weighted autoencoder decoder to form the metasurface image. Placing the same physics loss on top of both classical and quantum generators is what makes the comparison about the quantum generator rather than about the loss.

What would settle it

Run a full-wave finite-element simulation of the specific generated asymmetric unit cell reported to give $Q=1.56\times10^5$ at 10 µm, with the actual material dispersion, and compare the simulated absorption linewidth and peak to the target; if the simulated Q-factor falls below about $10^4$ or the resonance shifts, then the physics loss was fitting a surrogate that does not generalize to the high-Q regime.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the hybrid physics-informed quantum GAN solves the inverse metasurface design problem with extreme data efficiency. The generator is a variational quantum circuit that maps a nine-dimensional input (four Fano parameters plus five latent variables) through learned rotations and entangling gates to Pauli-Z expectation values; those are decoded by a pretrained importance-weighted autoencoder into a 64×64 RGB unit-cell image. The discriminator is dual-headed, classifying real versus fake while regressing the four Fano parameters, and the physics loss evaluates the Fano line-shape formula on those predictions. On the paper's reported numbers, this yields an MSE of $10^{-6}$ versus $10^{-3}$ for the plain GAN, a runtime of 1.26 h versus 25 h, and a training set of roughly 100 image–vector pairs rather than 20,000. The high-Q result is presented as emergent behavior: with a Q-factor floor $Q_{\min}=10^5$ added to the loss, the generator favors asymmetric geometries and produces simulated resonances with Q up to $1.56\times10^5$, although the training data contained only Q around $10^3$.

Load-bearing premise

The load-bearing premise is that the four Fano parameters the discriminator predicts from a generated image, which are the only inputs to the physics loss, agree with what a full electromagnetic simulation would produce for that geometry, especially for out-of-distribution asymmetric designs with Q-factors above $10^5$.

Editorial extensions

If this is right

Inverse design of narrow-band metasurface absorbers becomes a data-scarce problem: the physics-informed quantum GAN reportedly needs roughly 100 training image–vector pairs (64 in the conclusion) where the classical GAN baseline needs 20,000.
The reported 24× runtime reduction (1.26 h versus 25 h) and the MSE drop from about $10^{-3}$ to $10^{-6}$, if reproduced, make iterative design cycles feasible on one GPU instead of a long batch campaign.
With a Q-factor floor of $10^5$ in the loss, the generator shifts toward asymmetric unit cells and produces simulated Fano resonances with Q up to $1.56\times10^5$, which is the design route needed for quasi-bound-states-in-the-continuum sensors.
The fitted Fano parameters can be matched to real mid-infrared materials such as GaAs, so the output spectra can be tied to fabricable material systems rather than abstract refractive indices.
Because the physics loss is a spectral formula rather than a specific structure class, the same pipeline could be retargeted to other resonance lineshapes and other metasurface unit-cell families.

Reading between the lines

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

Editorial inference: The high-Q generation is an expected effect of optimization, since $L_Q$ explicitly enforces $Q_{\min}=10^5$; the open question is whether the discriminator's predicted Fano parameters for these out-of-distribution asymmetric designs match full-wave simulation.
Editorial inference: The '64 samples' figure is only an end-to-end sample count if the pretrained autoencoder's own training data is not a large metasurface dataset; auditing that pretraining set would settle whether the advantage is in the quantum generator or in the latent-space prior.
Editorial inference: A direct ablation—same IWAE decoder, same 100 samples, but a classical latent generator of comparable size instead of the variational circuit—would isolate how much of the gain comes from the quantum generator versus the physics loss and the autoencoder.
Editorial inference: For sensing applications, the decisive test is not the fitted Q-factor but the wavelength shift of the resonance under a refractive-index perturbation; running that perturbation simulation on the generated high-Q designs would connect the paper's Q-factors to usable sensitivity.

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.

Desk Editor's Note

read the letter

Colleague,

The core idea—adding a Fano resonance constraint to a GAN's loss and implementing it with a VQC generator—is a reasonable thing to try. The paper is transparent about borrowing LaSt-QGAN and the Fano model, and the COMSOL verification workflow for generated designs is the right instinct. If the spectra in Figures 6–8 really come from COMSOL, the high-Q results are physical and worth a look. But the evidence as presented does not support the headline claims.

The comparison is uncontrolled. The QGAN is trained on 64–100 ring-geometry samples, the classical GAN on 3000 samples of seven different geometries, with different discriminators (MLP vs CNN), different epochs, and different data. The 99.5% sample saving and the order-of-magnitude MSE improvement are therefore not meaningful. Table 2 has no error bars, and the numbers are inconsistent (64 vs 100 samples; runtime comparisons across different training lengths).

The larger methodological problem is that the physics loss is computed from the discriminator's predicted Fano parameters, not from simulated spectra. The generator is trained to satisfy the discriminator's regression head, which has never seen a true Q > 10^3 example. The paper never reports a calibration between predicted and simulated Q for generated designs. That leaves open the possibility that the high-Q results in Figure 8, if they are COMSOL simulations, are real but incidental, or that the network is exploiting the surrogate's blind spot. The stress-test note worries about circularity; I think 'circular' is too strong if the final spectra are independently simulated, but the missing calibration makes the claim under-supported.

The high-Q generation is also explicitly forced: Q_min=10^5 is in the loss. Calling it an emergent discovery overstates what was done. The observation that asymmetry correlates with high Q is consistent with quasi-BIC physics, so the result is not surprising.

The citation pattern looks fine; the relevant quantum GAN and PINN literature is cited. The writing is rough in places, but the architecture and loss definitions are clear.

Who is this for? Someone working on quantum ML for photonic inverse design, but they should not cite the quantitative claims until the comparison is fixed. I would send this to peer review because the combination is novel and the problems are addressable: rerun with matched baselines, report repeated runs with error bars, and validate the surrogate's predicted Q against COMSOL for generated images. The idea deserves a rigorous test; this version doesn't provide it.

Referee Report

4 major / 3 minor

Summary. The paper proposes a hybrid inverse-design pipeline in which a classical cDCGAN or a quantum VQC-based generator (LaSt-QGAN with a pretrained IWAE decoder) is trained with a physics-informed loss derived from a Fano absorption model. The authors claim that the QGAN+PINN framework converges faster, needs 99.5% fewer training samples, and achieves an order-of-magnitude lower MSE than conventional GANs, and that it generates metasurface geometries with Q-factors above 10^5 despite training data with Q~10^3. The manuscript reports comparisons in Table 2 and Figures 5–8, with COMSOL simulation mentioned as the evaluation pipeline.

Significance. If the claims were supported, the framework would be a useful contribution to data-scarce inverse design and would demonstrate a concrete benefit of quantum generative models in nanophotonics. The manuscript has several strengths: it addresses a relevant problem, integrates a spectral lineshape model into the loss, and describes a reproducible workflow from image to COMSOL simulation. However, as submitted the headline results are not established because they depend on the same surrogate used in the loss, the baselines are not matched, and no error bars or independent verification are provided. The work would be significant after a major rework that adds true COMSOL-validated spectra and controlled comparisons.

major comments (4)

[Section 2.3, Eq. (2); Section 3, Figs. 6–8] The central claim that QGAN+PINN achieves Q-factors up to 1.56×10^5 is not supported by independent simulation. L_Q in Eq. (2) directly enforces Q_val ≥ Q_min = 10^5, where Q_val = ν0/Γ is computed from the discriminator's regression head's predicted linewidth, and the training spectra have Q~10^3, so the regression has never seen a genuine high-Q spectrum. The workflow in Section 3 mentions COMSOL simulation, but Figures 6–8 and Table 2 do not report COMSOL-verified spectra or a calibration between predicted and simulated Q for generated images; without this, the reported high-Q values are consistent with the generator exploiting the surrogate's out-of-distribution behavior rather than with physically realized resonances.
[Table 2; Section 3] The headline comparisons are not controlled: the cDCGAN baseline uses 20,000 images of seven geometries, the QGAN uses 64 or 100 images of different ring/polygon geometries, the generators differ (CNN versus VQC plus pretrained IWAE decoder), and training epochs differ (500 for GAN+PINN versus 250 for QGAN+PINN). Consequently, 'faster convergence' and '99.5% fewer training samples' conflate architecture, dataset, and pretraining, and cannot be attributed to physics-informed quantum enhancement. The paper needs matched experiments with identical data, architecture, and training budget before such claims can be made.
[Section 2.1; Section 3; Conclusion] The '64 training samples' claim omits the pretrained IWAE. The QGAN generator is a VQC whose outputs are decoded by a pretrained IWAE decoder, and the manuscript does not disclose the dataset used to pretrain the IWAE. If the IWAE was trained on the same or a larger metasurface dataset, the claimed data-efficiency advantage is not measured by the 64 QGAN samples. The manuscript also contradicts itself on this count: Section 3 states 'only 100 images' while the abstract and conclusion state 64, and Table 2 lists 100 under 'Training Samples'.
[Table 2] Table 2 reports MSE values only as powers of ten (10^-3, 10^-4, 10^-5, 10^-6) and lists a single runtime per method, with no error bars, repeated initializations, or statistical significance tests. Since the central performance comparison rests on these numbers, the uncertainty is unquantified and the claimed 'order of magnitude' improvement cannot be assessed.

minor comments (3)

[Section 3, first paragraph] The sentence 'using the physical parameters — refractive index (n) and thickness (t), predicted by the physics-informed network' is inconsistent with the methodology, where the network predicts Fano parameters (ν0, Γ, A0, q); please clarify how n and t are obtained.
[Section 2.2 and Section 2.3] Equation (1) uses lower-case q for the Fano asymmetry parameter, while Section 2.3 defines the quality factor Q; to avoid confusion, reserve distinct symbols and define all variables at first use.
[Multiple locations] The training-sample count is reported as '64' in the abstract and conclusion, '100 images' in Section 3, and '500 images' in the discussion of QGAN-only in Section 3; these inconsistencies should be reconciled.

Circularity Check

2 steps flagged · score 6.0 of 10

The high-Q (10^5) result is imposed by the Q_min=10^5 term in the physics loss, and the headline MSE is evaluated on the same Fano parameters that the loss minimizes; both central claims therefore reduce in part to the training objective.

fitted input called prediction [Abstract; Section 2.3 (Physics informed Losses, L_Q definition); Section 3 (Figure 8 discussion)]
"LQ = (max(Qmin − Q_val,0))^2. This loss term ensures that the generated metasurface has a Q-factor greater than Qmin. ... we observe high quality factor resonances when the physics-informed loss includes the quality factor term LQ, where the threshold Qmin is set to 10^5. ... the model is able to generate highly asymmetric metasurface structures with Q-factors exceeding 10^5."

Q_val in the loss is not a COMSOL-measured linewidth; it is the discriminator's predicted ν0/Γ for the generated image. Setting Qmin=10^5 and minimizing (Qmin−Q_val)^2 directly drives the surrogate predicted Q to at least 10^5. The paper's 'remarkable' claim that Q>10^5 is generated despite training Q≈10^3 is therefore the expected outcome of the explicitly imposed constraint, not an extrapolation discovered from data. Without an independent COMSOL simulation and calibration of the generated images' Q values, the reported Q>10^5 is the loss-optimized surrogate value by construction.
other [Section 2.3, Eq. (2); Section 3 (Table 2 and evaluation description)]
"Lphysics =λA ·LA +λν0 ·Lν0 +λQ ·LQ (2) ... Unlike previous works that compute mean squared error (MSE) over the entire absorption spectrum, our approach focuses on key resonance characteristics, namely, the peak absorption, target resonance frequency, and linewidth for evaluation and comparison."

The MSE used in Table 2 is computed on the same key quantities that appear in the physics loss: peak absorption is penalized by LA, resonance frequency by Lν0, and linewidth-derived Q by LQ. Thus the reported order-of-magnitude MSE advantage of QGAN+PINN is partly a measure of how well the model minimized its own objective, not an independent spectral-fidelity benchmark. The relative comparison between models retains some meaning, but the absolute MSE values reduce to the training objective unless the spectra are independently simulated and fitted.

full rationale

The paper contains no load-bearing self-citation loop and no imported uniqueness theorem; the cited prior works are external and are not used to forbid alternative frameworks. The circularity burden comes from the loss/evaluation design. First, the high-Q claim: the abstract presents Q>10^5 as a remarkable emergent capability, but Section 2.3 defines LQ = (max(Qmin − Q_val,0))^2 with Qmin=10^5, where Q_val is the discriminator's predicted linewidth ratio. Minimizing this loss makes the surrogate Q reach the threshold by construction; calling the resulting value a generation capability is reporting the constraint rather than an independent discovery. Second, the headline MSE comparisons are 'focused on key resonance characteristics' that are exactly the targets of the physics loss, so the absolute MSE values are not an external test of spectral accuracy. A COMSOL verification of the generated geometries would break the circularity, and the paper describes a COMSOL workflow, but it does not report per-figure COMSOL calibration of the surrogate-predicted Q or clearly separate which spectra in Figures 6-8 and Table 2 come from independent simulation. These issues make the central performance claims partially self-referential, though not fully definitionally equivalent; hence a score of 6.

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

The central claims rest on four unproved inputs: the Fano line shape as universal model, an unvalidated surrogate predictor, an undisclosed pretrained autoencoder, and trust in COMSOL as ground truth. The model also has explicit free parameters, especially the high-Q threshold that manufactures the headline result.

free parameters (5)

lambda_A = 26
Bayesian optimization selected this weight for the absorption loss; directly affects reported MSE.
lambda_nu0 = 14
Resonance frequency loss weight tuned by Bayesian optimization.
lambda_Q = 26
Q-factor loss weight tuned by Bayesian optimization; controls strength of the high-Q constraint.
Q_min = 10^5
Hand-chosen threshold in L_Q; the paper's high-Q results match this imposed threshold.
A_target = 0.9
Target absorption in L_A; chosen by the authors as the desired peak absorption.

assumptions (4)

domain assumption Fano profile A(nu)=A0*(q+delta)^2/(1+delta^2) models the absorption of every generated metasurface.
Section 2.2 uses this model for the loss and parameter fitting; not established for all arbitrary generated geometries.
ad hoc to paper Discriminator-predicted Fano parameters are a valid forward surrogate for COMSOL-simulated absorption spectra.
Physics loss (Eq. 2) uses predicted parameters rather than simulated spectra; no accuracy validation on generated out-of-distribution images is reported.
ad hoc to paper Pretrained IWAE preserves the information needed to reconstruct physical metasurface images from latent codes.
Section 2.1 relies on IWAE for decoding VQC outputs; its pretraining data is not disclosed, and its reconstruction fidelity is not quantified.
standard math COMSOL FEM simulations provide ground-truth absorption spectra.
Used as evaluation reference throughout Section 3; reasonable but no experimental verification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Inverse Design using Physics-Informed Quantum GANs for Tailored Absorption in Dielectric Metasurfaces." pith.science (2026). https://pith.science/paper/4MHZGB4X

@misc{pith2026250718132,
  author       = {Pith},
  title        = {Pith review of: Inverse Design using Physics-Informed Quantum GANs for Tailored Absorption in Dielectric Metasurfaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MHZGB4X}},
  note         = {Machine review of arXiv:2507.18132}
}

abstract

High Q-factor narrow-band absorption exhibits high spectral selectivity enabling high-sensitive photodetectors, sensors and thermal emitters. All-dielectric metasurfaces are widely regarded as excellent candidates for giving rise to such narrow-band absorption. However, designing metasurfaces with specific functionalities remains a challenging task both experimentally and computationally, which is why inverse design methods are increasingly being explored. Inverse design process is highly complex due to its non-unique solutions and the higher dimensionality of the design space, making it challenging to precisely control the resonance wavelength, linewidth, and absorption intensity. In this paper, we present a novel hybrid methodology that integrates generative adversarial networks (GANs) (both classical and quantum) with physics-informed neural networks (PINNs) for the inverse design of narrow-band absorbing metasurfaces. By introducing a Fano-shaped absorption spectrum equation into the PINN loss function, we enforce physical constraints on the resonance behavior, ensuring outputs that are both spectrally accurate and physically consistent. The study presents a comparison between a conventional GAN + PINN framework and a PINN augmented by a hybrid quantum-classical GAN (QGAN). The findings indicate that the integrated PINN + QGAN model achieves faster convergence, requires 99.5\% fewer training samples, and yields an order of magnitude lower MSE compared to conventional GANs. Remarkably, even though the training dataset only contains metasurfaces with Q-factors on the order of $10^3$, the model is able to generate highly asymmetric metasurface structures with Q-factors exceeding $10^5$. This study presents a novel framework that integrates quantum machine learning with physics-based modeling, providing a promising method for quantum-enhanced inverse design in nanophotonic systems.

Figures

Figures reproduced from arXiv: 2507.18132 by the authors.

**Figure 2.** a) Schematic of the quantum-enhanced QGAN + PINN framework. b) Variational [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Flowchart of the iterative optimization process for determining optimal Fano [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Workflow for computing the absorption spectrum from GAN-generated metasur [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

**Figure 5.** Figure 5: Inverse design outcomes from different generative models for a sharp Fano absorp [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: Comparison of absorption spectra generated by different generative models—GAN, [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

**Figure 7.** Figure 7: Comparison of absorption spectra for a) GAN + PINN and b) QGAN + PINN [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

**Figure 8.** Figure 8: Absorption spectra of metasurfaces exhibiting high-Q Fano resonances at target [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages

[1]

F.; Jokerst, N

Liu, X.; Tyler, T.; Starr, T.; Starr, A. F.; Jokerst, N. M.; Padilla, W. J. Taming the Blackbody with Infrared Metamaterials as Selective Thermal Emitters. Phys. Rev. Lett. 2011, 107, 045901

work page 2011
[2]

Metasurface Photodetectors

Li, J.; Li, J.; Zhou, S.; Yi, F. Metasurface Photodetectors. Micromachines 2021, 12, 1584

work page 2021
[3]

Multiband-switchability and high-absorptivity of a metamaterial perfect absorber based on a plasmonic resonant structure in the near-infrared region

Liang, J.; Chen, Y.; Zhou, Z.; Chen, S. Multiband-switchability and high-absorptivity of a metamaterial perfect absorber based on a plasmonic resonant structure in the near-infrared region. RSC Advances 2022, 12, 30871--30878

work page 2022
[4]

Terahertz narrowband perfect metasurface absorber based on micro-ring-shaped GaAs array for enhanced refractive index sensing

Cheng, Y.; Qian, Y.; Luo, H.; Chen, F.; Cheng, Z. Terahertz narrowband perfect metasurface absorber based on micro-ring-shaped GaAs array for enhanced refractive index sensing. Physica E: Low-dimensional Systems and Nanostructures 2023, 146, 115527

work page 2023
[5]

Ultra-Narrow Band Mid-Infrared Perfect Absorber Based on Hybrid Dielectric Metasurface

Chen, S.; Chen, Z.; Liu, J.; Cheng, J.; Zhou, Y.; Xiao, L.; Chen, K. Ultra-Narrow Band Mid-Infrared Perfect Absorber Based on Hybrid Dielectric Metasurface. Nanomaterials 2019, 9, 1350

work page 2019
[6]

W.; Raman, A

Yeung, C.; Tsai, R.; Pham, B.; King, B.; Kawagoe, Y.; Ho, D.; Liang, J.; Knight, M. W.; Raman, A. P. Global Inverse Design across Multiple Photonic Structure Classes Using Generative Deep Learning. Advanced Optical Materials 2021, 9, 2100548

work page 2021
[7]

Raissi, M.; Perdikaris, P.; Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 2019, 378, 686--707

work page 2019
[8]

E.; Dal Negro, L

Chen, Y.; Lu, L.; Karniadakis, G. E.; Dal Negro, L. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Optics Express 2020, 28, 11618--11633

work page 2020

Show all 23 references

[9]

Physics-Based Approach for a Neural Networks Enabled Design of All-Dielectric Metasurfaces

Tanriover, I.; Hadibrata, W.; Aydin, K. Physics-Based Approach for a Neural Networks Enabled Design of All-Dielectric Metasurfaces. ACS Photonics 2020, 7, 1957--1964

2020
[10]

J.; Jin, B.; Chen, J.; Wu, P

Xu, Y.; Yang, J.-Q.; Fan, K.; Wang, S.; Wu, J.; Zhang, C.; Zhan, D.-C.; Padilla, W. J.; Jin, B.; Chen, J.; Wu, P. Physics-Informed Inverse Design of Programmable Metasurfaces. Advanced Science 2024, 11, 2406878

2024
[11]

Quantum Computing in the NISQ Era and Beyond

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

2018
[12]

Quantum generative adversarial learning

Lloyd, S.; Weedbrook, C. Quantum generative adversarial learning. Physical Review Letters 2018, 121, 040502

2018
[13]

Quantum generative adversarial networks

Dallaire-Demers, P.-L.; Killoran, N. Quantum generative adversarial networks. Physical Review A 2018, 98, 012324

2018
[14]

Huang, H.-L. et al. Experimental Quantum Generative Adversarial Networks for Image Generation . Phys. Rev. Appl. 2021, 16, 024051

2021
[15]

L.; West, M

Tsang, S. L.; West, M. T.; Erfani, S. M.; Usman, M. Hybrid Quantum–Classical Generative Adversarial Network for High-Resolution Image Generation. IEEE Transactions on Quantum Engineering 2023, 4, 1--19

2023
[16]

A.; Baheri, B.; Chen, D.; Mao, Y.; Guan, Q.; Li, A.; Fang, B.; Xu, S

Stein, S. A.; Baheri, B.; Chen, D.; Mao, Y.; Guan, Q.; Li, A.; Fang, B.; Xu, S. QuGAN: A Quantum State Fidelity based Generative Adversarial Network. 2021 IEEE International Conference on Quantum Computing and Engineering (QCE). 2021; pp 71--81

2021
[17]

Variational Quantum Generators: Generative Adversarial Quantum Machine Learning for Continuous Distributions

Romero, J.; Aspuru-Guzik, A. Variational Quantum Generators: Generative Adversarial Quantum Machine Learning for Continuous Distributions. Advanced Quantum Technologies 2021, 4, 2000003

2021
[18]

Y.; Thanasilp, S.; Saux, B

Chang, S. Y.; Thanasilp, S.; Saux, B. L.; Vallecorsa, S.; Grossie, M. Latent Style-based Quantum GAN for High-quality Image Generation. 2024; P reprint at https://arxiv.org/abs/2406.02668

2024 arXiv
[19]

Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Radford, A.; Metz, L.; Chintala, S. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. 2015; P reprint at https://arxiv.org/abs/1511.06434

2015 arXiv
[20]

2018; P reprint at https://arxiv.org/abs/1811.04968

Bergholm, V.; others PennyLane: Automatic differentiation of hybrid quantum-classical computations. 2018; P reprint at https://arxiv.org/abs/1811.04968

2018 arXiv
[21]

Bayesian Approach to Global Optimization: Theory and Applications; Mathematics and its Applications; Springer Dordrecht, 1989

Mockus, J. Bayesian Approach to Global Optimization: Theory and Applications; Mathematics and its Applications; Springer Dordrecht, 1989

1989
[22]

Conditional Generative Adversarial Nets

Mirza, M.; Osindero, S. Conditional Generative Adversarial Nets. 2014; P reprint at https://arxiv.org/abs/1411.1784

2014 arXiv
[23]

H ! B ! B B B ! B ! ! rB ! B ! B 9! B ! B !B B ! B ! !DBN ! B ! B

Ciftci, K.; Hackl, K. A physics-informed GAN framework based on model-free data-driven computational mechanics. Computer Methods in Applied Mechanics and Engineering 2024, 424, 116907 mcitethebibliography main.tex0000664000000000000000000012146515040353150011231 0ustar rootroo...

2024

Pith tools

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

[1] [1]

F.; Jokerst, N

Liu, X.; Tyler, T.; Starr, T.; Starr, A. F.; Jokerst, N. M.; Padilla, W. J. Taming the Blackbody with Infrared Metamaterials as Selective Thermal Emitters. Phys. Rev. Lett. 2011, 107, 045901

work page 2011

[2] [2]

Metasurface Photodetectors

Li, J.; Li, J.; Zhou, S.; Yi, F. Metasurface Photodetectors. Micromachines 2021, 12, 1584

work page 2021

[3] [3]

Multiband-switchability and high-absorptivity of a metamaterial perfect absorber based on a plasmonic resonant structure in the near-infrared region

Liang, J.; Chen, Y.; Zhou, Z.; Chen, S. Multiband-switchability and high-absorptivity of a metamaterial perfect absorber based on a plasmonic resonant structure in the near-infrared region. RSC Advances 2022, 12, 30871--30878

work page 2022

[4] [4]

Terahertz narrowband perfect metasurface absorber based on micro-ring-shaped GaAs array for enhanced refractive index sensing

Cheng, Y.; Qian, Y.; Luo, H.; Chen, F.; Cheng, Z. Terahertz narrowband perfect metasurface absorber based on micro-ring-shaped GaAs array for enhanced refractive index sensing. Physica E: Low-dimensional Systems and Nanostructures 2023, 146, 115527

work page 2023

[5] [5]

Ultra-Narrow Band Mid-Infrared Perfect Absorber Based on Hybrid Dielectric Metasurface

Chen, S.; Chen, Z.; Liu, J.; Cheng, J.; Zhou, Y.; Xiao, L.; Chen, K. Ultra-Narrow Band Mid-Infrared Perfect Absorber Based on Hybrid Dielectric Metasurface. Nanomaterials 2019, 9, 1350

work page 2019

[6] [6]

W.; Raman, A

Yeung, C.; Tsai, R.; Pham, B.; King, B.; Kawagoe, Y.; Ho, D.; Liang, J.; Knight, M. W.; Raman, A. P. Global Inverse Design across Multiple Photonic Structure Classes Using Generative Deep Learning. Advanced Optical Materials 2021, 9, 2100548

work page 2021

[7] [7]

Raissi, M.; Perdikaris, P.; Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 2019, 378, 686--707

work page 2019

[8] [8]

E.; Dal Negro, L

Chen, Y.; Lu, L.; Karniadakis, G. E.; Dal Negro, L. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Optics Express 2020, 28, 11618--11633

work page 2020