arxiv: 2605.03335 · v1 · submitted 2026-05-05 · ⚛️ physics.app-ph

Recognition: unknown

WGAN based Inverse Design of Active Dual Band FSS with Switchable Transmission

Huanran Qiu, Long Li, Rui Xi, Shiyun Ma, Xiaokui Kang, Xinke Kuang, Ying Li, Yuanyuan Wang

Pith reviewed 2026-05-07 12:40 UTC · model grok-4.3

classification ⚛️ physics.app-ph

keywords inverse designfrequency selective surfaceWGANactive FSSdual-bandpin diodestopology generationelectromagnetic response

0 comments

The pith

A WGAN learns to map target electromagnetic responses directly to fabricable topologies for active dual-band FSS with pin-diode switching.

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

The paper sets out to create an inverse design workflow for a switchable dual-band frequency selective surface that transmits at high frequencies but reflects at low frequencies when pin diodes change state. It introduces a crystal growth method to produce candidate topologies and trains a simplified U-Net Wasserstein GAN to turn a desired electromagnetic response into the corresponding structure parameters. A reader would care if this holds because conventional FSS design relies on slow manual tuning that restricts how complex the switching behavior can become. The model reports 95.59 percent training and 90.84 percent validation accuracy on the WGAN task, with the U-Net reaching 98.5 percent and 94.1 percent, and the outputs are checked against full-wave simulations plus physical measurements.

Core claim

The authors establish that a crystal growth-based topology generator combined with a simplified U-Net Wasserstein GAN can learn an inverse mapping from specified electromagnetic response to structure topology parameters, enabling the creation of an active dual-band FSS whose high-frequency passband remains stable while its low-frequency behavior switches from transmission to reflection according to pin-diode state. The trained model reaches 95.59 percent training accuracy and 90.84 percent validation accuracy for the WGAN component and 98.5 percent training and 94.1 percent validation for the U-Net, after which the generated topologies are shown to satisfy the target responses in full-wave电磁

What carries the argument

The simplified U-Net Wasserstein GAN model that learns the inverse mapping from electromagnetic response to structure topology parameters.

If this is right

Designers can specify a desired frequency response and receive candidate topologies without repeated manual optimization loops.
The resulting structures can be fabricated and will exhibit the intended high-frequency transmission together with diode-controlled low-frequency switching.
The approach reduces the time required for exploring complex active FSS geometries compared with conventional trial-and-error methods.
Design flexibility increases because the inverse model supports direct specification of both passband and switching characteristics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same inverse-mapping strategy could be tested on other reconfigurable surfaces whose behavior depends on different control elements such as varactors or MEMS switches.
Adding fabrication-tolerance constraints directly into the topology generator might further close the gap between simulated and measured performance.
If the accuracy holds across wider parameter ranges, the method points toward automated pipelines that generate entire families of related FSS designs from a single specification.

Load-bearing premise

The mapping learned from the training set will continue to produce physically realizable topologies whose measured electromagnetic behavior with added pin diodes matches the target response.

What would settle it

Fabricate one or more generated topologies, integrate the pin diodes, and measure the actual transmission and reflection spectra in both diode states to see whether they deviate beyond acceptable error from the specified dual-band switchable targets.

read the original abstract

This letter presents a novel design method for switchable dual band transmissive frequency selective surface (FSS). The proposed FSS possesses characteristics of maintaining passband characteristics at high frequencies, while switching from transmission to reflection at low frequencies with pin diodes states altering. Specifically, we propose a crystal growth-based topology generation strategy, and utilize a simplified U-Net Wasserstein GAN (WGAN) neural network model to establish an inverse mapping model from electromagnetic response to structure topology parameters. The trained WGAN achieves training and validation accuracies of 95.59% and 90.84%, while the simplified U-Net attains training and validation accuracies of 98.5% and 94.1%. Using the trained WGAN. The generated structural topologies were validated through full-wave simulations and experimental measurements. The proposed method enhances the design flexibility and overcomes the time- consuming drawbacks of conventional FSS design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a workable ML pipeline for inverse design of pin-diode switchable dual-band FSS using crystal-growth topologies and a simplified U-Net WGAN, with simulation-plus-measurement checks, but the end-to-end match after diode insertion rests on internal validation numbers whose exact meaning is not fully clear from the abstract.

read the letter

The main takeaway is that this work supplies a concrete generative-model route for designing reconfigurable FSS that switch from transmission to reflection at one band while keeping a passband at the other. The authors combine a crystal-growth topology generator with a simplified U-Net WGAN to learn the inverse map from target EM response to geometry parameters, then show the outputs through full-wave simulation and fabricated prototypes. That is a practical step for people who design active metasurfaces for wireless systems and want to cut down on manual iteration.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a crystal growth-based topology generation strategy combined with a simplified U-Net Wasserstein GAN (WGAN) to perform inverse design of an active dual-band transmissive frequency selective surface (FSS) whose low-frequency transmission can be switched to reflection via pin-diode bias states. The paper reports training/validation accuracies of 95.59 % / 90.84 % for the WGAN and 98.5 % / 94.1 % for the U-Net, states that the generated topologies were validated by full-wave simulation and experimental measurement, and claims the approach overcomes the time-consuming nature of conventional FSS design.

Significance. If the inverse mapping reliably yields fabricable topologies whose measured S-parameters match the design targets after explicit diode integration, the method would meaningfully accelerate exploration of switchable FSS geometries. The use of a generative model to map response to topology parameters is a relevant direction for electromagnetic inverse design; however, the reported internal accuracies do not yet demonstrate end-to-end physical fidelity.

major comments (3)

[Abstract] Abstract: the reported accuracies (WGAN 95.59 % / 90.84 %, U-Net 98.5 % / 94.1 %) are presented without definition of the underlying metric (pixel overlap, parameter RMSE, or S-parameter fidelity). Because the central claim is that the learned inverse mapping produces structures whose electromagnetic response matches the target after diode insertion, the absence of this definition prevents evaluation of whether the numbers support the claim.
[Abstract] Abstract: no information is given on training-set size, validation-set size, data-generation pipeline, or whether the validation structures include the same crystal-growth rules and diode-placement assumptions used for training. This leaves open the possibility that the 90.84 % validation accuracy reflects interpolation within the simulation assumptions rather than generalization to physically realizable devices.
[Abstract] Abstract: the statement that 'generated structural topologies were validated through full-wave simulations and experimental measurements' is not accompanied by quantitative error metrics, S-parameter comparison plots, or discussion of bias-line parasitics and diode package effects. These elements lie outside the training distribution and are load-bearing for the claim that the inverse design remains accurate once the structure is fabricated and biased.

minor comments (1)

[Abstract] The abstract refers to a 'simplified U-Net' without stating which architectural modifications were made relative to the standard U-Net; a brief description would improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help improve the clarity of our manuscript. We address each major comment point by point below and will revise the abstract accordingly to incorporate the requested definitions and details.

read point-by-point responses

Referee: [Abstract] Abstract: the reported accuracies (WGAN 95.59 % / 90.84 %, U-Net 98.5 % / 94.1 %) are presented without definition of the underlying metric (pixel overlap, parameter RMSE, or S-parameter fidelity). Because the central claim is that the learned inverse mapping produces structures whose electromagnetic response matches the target after diode insertion, the absence of this definition prevents evaluation of whether the numbers support the claim.

Authors: We agree that the accuracy metric requires explicit definition in the abstract. The reported figures represent pixel-overlap accuracy between the generated binary topology maps and the ground-truth structures produced by the crystal-growth algorithm. This metric quantifies the fidelity of the inverse mapping from response to topology. We will revise the abstract to state this definition clearly. revision: yes
Referee: [Abstract] Abstract: no information is given on training-set size, validation-set size, data-generation pipeline, or whether the validation structures include the same crystal-growth rules and diode-placement assumptions used for training. This leaves open the possibility that the 90.84 % validation accuracy reflects interpolation within the simulation assumptions rather than generalization to physically realizable devices.

Authors: The training and validation sets were generated using the identical crystal-growth topology strategy and diode-placement rules described in Section II. We will add the dataset sizes and a statement confirming that validation samples follow the same generation rules to the revised abstract, while the full data-generation pipeline remains detailed in the main text. revision: yes
Referee: [Abstract] Abstract: the statement that 'generated structural topologies were validated through full-wave simulations and experimental measurements' is not accompanied by quantitative error metrics, S-parameter comparison plots, or discussion of bias-line parasitics and diode package effects. These elements lie outside the training distribution and are load-bearing for the claim that the inverse design remains accurate once the structure is fabricated and biased.

Authors: We acknowledge that the abstract would benefit from quantitative support for the validation claim. The main manuscript provides S-parameter comparison plots and error metrics from both full-wave simulations and measurements (including discussion of bias-line parasitics and diode effects in Section IV). We will revise the abstract to reference these quantitative results and note that parasitic effects were accounted for in the experimental validation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; validation includes external full-wave and measurement benchmarks

full rationale

The paper trains a WGAN and U-Net on paired simulation data (topology parameters ↔ EM responses) generated by a crystal-growth strategy and full-wave solver. It reports training/validation accuracies on held-out data from the same pipeline (standard supervised learning practice) and then explicitly validates generated topologies via additional full-wave simulations plus physical fabrication and measurement. No step reduces a claimed prediction or result to its inputs by definition, self-citation, or renaming; the end-to-end claim rests on the external measurement agreement rather than internal fit statistics alone. This matches the most common honest non-finding for ML inverse-design papers.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the learned inverse mapping being accurate and generalizable. Training the GAN introduces many fitted parameters whose values are not reported. The mapping itself is treated as a domain assumption rather than derived from first principles.

free parameters (2)

WGAN and U-Net hyperparameters
Architecture depth, learning rate, batch size, and loss weights are chosen to reach the stated accuracies but are not enumerated.
Crystal-growth topology parameters
Rules controlling how candidate structures are generated before the GAN sees them.

axioms (2)

domain assumption Electromagnetic response is a deterministic function of geometry and diode state that can be inverted by a neural network.
Core premise enabling the inverse-design claim; invoked when the WGAN is trained to map response back to topology.
domain assumption Full-wave simulation results are sufficiently accurate proxies for physical measurements.
Used to generate training data and to claim validation.

pith-pipeline@v0.9.0 · 5472 in / 1641 out tokens · 69287 ms · 2026-05-07T12:40:40.822665+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

[1]

Dual-band multi-bit programmable reflective metasurface unit cell: design and experiment,

Y. Saifullah et al., “Dual-band multi-bit programmable reflective metasurface unit cell: design and experiment,” Opt. Express., vol. 29, no. 2, pp. 2658–2668, Jan 2021, doi: 10.1364/OE.415730

work page doi:10.1364/oe.415730 2021
[2]

Dual-band and dual-polarized programmable metasurface unit with independent channels

N., L. Yun et al., “Dual-band and dual-polarized programmable metasurface unit with independent channels.” J. Phys. D: Appl. Phys., vol. 54, no. 14, p. 145109, Jan 2021, doi: 10.1088/1361-6463/abd713

work page doi:10.1088/1361-6463/abd713 2021
[3]

A Dual-Broadband Liquid-Crystal Programmable Metasurface and Its Application in Terahertz Wireless Communications

Y. Fu, Y. Li et al., “A Dual-Broadband Liquid-Crystal Programmable Metasurface and Its Application in Terahertz Wireless Communications.” Engineering, 2025, doi: 10.1016/j.eng.2025.08.040

work page doi:10.1016/j.eng.2025.08.040 2025
[4]

A dual-band frequency-selective surface with rectangular filtering characteristic for the second band enabled by harmonics resonance

J. Chen, X Luo et al., “A dual-band frequency-selective surface with rectangular filtering characteristic for the second band enabled by harmonics resonance.” AEU-Int. J. Electron. Commun., vol. 202, Dec 2025, Art. no. 155961, doi: 10.1016/j.aeue.2025.155961

work page doi:10.1016/j.aeue.2025.155961 2025
[5]

A programmable-metasurface-based TDMA fast beam switching communication system at 28 GHz

Y. Li, J. Eisenbeis et al., “A programmable-metasurface-based TDMA fast beam switching communication system at 28 GHz.” IEEE Antennas Wireless Propag. Lett., vol. 20, no. 5, pp. 658–662, May 2021, doi: 10.1109/LAWP.2021.3059322

work page doi:10.1109/lawp.2021.3059322 2021
[6]

Radar one-dimensional range profile dynamic jamming based on programmable metasurface

H. Xu et al., “Radar one-dimensional range profile dynamic jamming based on programmable metasurface.” IEEE Antennas Wireless Propag. Lett., vol. 20, no. 10, pp. 1883-1887, Oct 2021, doi: 10.1109/LAWP.2021.3098698

work page doi:10.1109/lawp.2021.3098698 2021
[7]

Low Sidelobe Dual-Beam Metasurface Antenna Based on Taylor Distribution and Digital Filtering Methods

Q. Liu et al., “Low Sidelobe Dual-Beam Metasurface Antenna Based on Taylor Distribution and Digital Filtering Methods.” IEEE Antennas Wireless Propag. Lett., vol. 24, no. 4, pp. 1003-1007, Apr 2025, doi: 10.1109/LAWP.2024.3523945

work page doi:10.1109/lawp.2024.3523945 2025
[8]

Phase-to-pattern inverse design for a fast realization of a functional metasurface by combining a deep neural network and a genetic algorithm

G. Wu et al., “Phase-to-pattern inverse design for a fast realization of a functional metasurface by combining a deep neural network and a genetic algorithm.” Opt. Express., vol. 30, no. 25, pp. 45612-45613, Dec 2022, doi: 10.1364/OE.478084

work page doi:10.1364/oe.478084 2022
[9]

Genetic algorithm assisted meta-atom design for high- performance metasurface optics

Z. Yu et al., “Genetic algorithm assisted meta-atom design for high- performance metasurface optics.” Opto-Electron Sci., vol. 3, no. 9, p. 240016, Sep 2024, doi: 10.29026/oes.2024.240016

work page doi:10.29026/oes.2024.240016 2024
[10]

Designing broadband cross-polarization conversion metasurfaces using binary particle swarm optimization algorithm

J. Wang et al., “Designing broadband cross-polarization conversion metasurfaces using binary particle swarm optimization algorithm.” Mater. Des., vol. 247, Nov 2024, Art. no. 113419, doi: 10.1016/j.matdes.2024.113419

work page doi:10.1016/j.matdes.2024.113419 2024
[11]

Particle Swarm Optimization for the Design of Frequency Selective Surfaces

S. Genovesi et al., “Particle Swarm Optimization for the Design of Frequency Selective Surfaces.” IEEE Antennas Wireless Propag. Lett., vol. 5, pp. 277-279, Dec 2006, doi: 10.1109/LAWP.2006.875900

work page doi:10.1109/lawp.2006.875900 2006
[12]

Numerical Optimization Methods for Metasurfaces

M.M.R. Elsawy et al., “Numerical Optimization Methods for Metasurfaces.” Laser Photon. Rev., vol. 14, Aug 2020, Art. no. 1900445, doi: 10.1002/lpor.201900445

work page doi:10.1002/lpor.201900445 2020
[13]

Review of numerical optimization techniques for meta-device design [Invited]

S.D. Campbell et al., “Review of numerical optimization techniques for meta-device design [Invited].” Opt. Mater. Express., vol. 9, no. 4, pp. 1842-1863, 2019, doi: 10.1364/OME.9.001842

work page doi:10.1364/ome.9.001842 2019
[14]

Deep convolutional generative adversarial networks assisted inverse design of quad-channel full-space metasurface

X. Liu et al., “Deep convolutional generative adversarial networks assisted inverse design of quad-channel full-space metasurface.” Opt. Express., vol. 32, no. 26, pp.46542-46556, 2024, doi: 10.1364/OE.541287

work page doi:10.1364/oe.541287 2024
[15]

Conditional Generative Adversarial Networks for Inverse Design of Multifunctional Metasurfaces

M. Kiani et al., “Conditional Generative Adversarial Networks for Inverse Design of Multifunctional Metasurfaces.” Adv. Photonics Res., vol. 3, no. 11, Aug 2022, Art. no. 2200110, doi: 10.1002/adpr.202200110

work page doi:10.1002/adpr.202200110 2022
[16]

Inverse Design of Bifunctional Metasurfaces Using Improved Generative Adversarial Networks

X. Liu et al., “Inverse Design of Bifunctional Metasurfaces Using Improved Generative Adversarial Networks.” IEEE Antennas Wireless Propag. Lett.,vol. 24, no. 3, pp. 582-586, Mar 2025, doi: 10.1109/LAWP.2024.3508094

work page doi:10.1109/lawp.2024.3508094 2025
[17]

Improved Training of Wasserstein GANs

I. Gulrajani et al., “Improved Training of Wasserstein GANs.” Adv. Neural Inf. Process. Syst., vol. 30, 2017

2017
[18]

Transmissive Metasurface Synthesis From Far-Field Masks Using Unsupervised Learning

C. Niu et al., “Transmissive Metasurface Synthesis From Far-Field Masks Using Unsupervised Learning.” IEEE Antennas Wireless Propag. Lett., Vol. 23, no. 8, pp. 2371-2375, Apr 2024, doi: 10.1109/LAWP.2024.3392158

work page doi:10.1109/lawp.2024.3392158 2024
[19]

A Solution to the Dilemma for FSS Inverse Design Using Generative Models

Z. Gu et al., "A Solution to the Dilemma for FSS Inverse Design Using Generative Models." IEEE Trans Antennas Propag., vol. 71, no. 6, pp. 5100-5109, June 2023, doi: 10.1109/TAP.2023.3266053

work page doi:10.1109/tap.2023.3266053 2023