Source-linked AI summary
Multifunctional Metasurface Design with a Generative Adversarial Network
Sensong An, Bowen Zheng, Hong Tang, Mikhail Y. Shalaginov, Li Zhou, Hang Li, Tian Gu, Juejun Hu, Clayton Fowler, Hualiang Zhang
TL;DR
The paper addresses instability and inapplicability in training a meta-atom discriminator for real-sample distributions. It introduces geometry interpolation to support distribution exploration, producing diverse free-form patterns and transmission-spectrum predictions over 30–60 THz, with designs evaluated under specified geometric conditions.
Problem
Conventional numerical interpolation is not applicable for the meta-atom discriminator, and stable Wasserstein distance results for real samples are difficult to obtain during training.
Method
The method combines random geometry portions through a novel geometry interpolation approach that allows the generator to extrapolate and explore the ground-truth distribution.
Results
The PNN precisely predicts transmission spectra for free-form meta-atom designs within the 30 to 60 THz frequency range, while generated patterns show more diverse shapes and refined details.
Takeaways & Limitations
Geometry interpolation supports generator exploration beyond the ground-truth distribution while retaining diverse structural details in the generated patterns.
Takeaways & Limitations
The training data use a low-refractive-index setting with n2 = 1.4, a 2.8 × 2.8 μm2 unit cell, and 0.4 μm minimum adjacent spacing.
Abstract
from arXiv · showhide
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.
Supporting Information
The paper presents a multifunctional metasurface design approach using a generative adversarial network.
- The paper is titled “Multifunctional Metasurface Design with a Generative Adversarial Network.”
- The listed authors include Sensong An, Bowen Zheng, Hong Tang, and collaborators.
1. Detailed network architecture
The proposed architecture combines a GAN generator and discriminator with predicting neural networks that evaluate generated meta-atom responses. The PNN predicts transmission spectra efficiently over 30–60 THz, supporting faster design evaluation.
- Network architecture: Input design conditions, including frequency-dependent amplitude and phase responses, polarization dependence, and material states, are combined with random noise.
- Network architecture: The GAN combines a generator that produces meta-atom images with a discriminator trained through Wasserstein distance.The generator uses transposed convolutions to form 2D designs, while the discriminator processes inputs through convolutional layers.
- Network architecture: The generator and discriminator are trained alternately, with the generator updated using the current discriminator.
- Predicting neural network: Two PNNs predict the real and imaginary transmission-spectrum components, from which transmissive amplitude and phase are derived.
- Predicting neural network: 30 to 60 THz is the frequency range in which the PNN predicts transmission spectra for free-form meta-atom designs.Unlike full-wave simulation tools, the PNN characterizes each meta-atom on a one-time calculation basis and thereby speeds the design process.
2. Training data collection
Training data consist of simulated, symmetric all-dielectric meta-atoms generated from randomly placed rectangular bars. The collection process produced 69,000 simulated structures, with 29,000 retained after removing similar patterns.
- Structure and materials: The all-dielectric meta-atom uses a 1 μm-thick high-index component on a low-index substrate with a 2.8 × 2.8 μm2 unit cell.The substrate refractive index is n2 = 1.4, while the high-index component is preferably n1 = 5.
- Pattern generation: The “needle drop” method randomly places 3 to 7 rectangular bars with 0.1 μm minimum resolution inside the square lattice.
- Pattern generation: A minimum spacing of 0.4 μm separates adjacent meta-atoms to minimize inter-cell coupling.
- Pattern generation: Patterns generated in the top-left quadrant are mirrored along the x and y axes to form the complete meta-atom.
- Simulation dataset: 69,000 meta-atoms with different shapes were simulated, and 29,000 structures were retained after similar patterns were removed.Data collection used eight parallel servers and was completed in 3 days.
3. Customized gradient-penalty method
The method replaces conventional numerical interpolation in WGAN-GP with geometry interpolation tailored to binary meta-atom patterns. This preserves physically meaningful intermediate samples and supports exploration beyond incompletely sampled training data.
- Wasserstein constraint: WGAN-GP uses a gradient penalty to enforce a discriminator gradient norm of 1 almost everywhere and thereby satisfy the 1-Lipschitz constraint.
- Motivation: Conventional interpolation can produce values between 0 and 1 that do not correspond to physical meta-atom structures.This makes stable Wasserstein-distance evaluation difficult for real samples during training.
- Customized interpolation: The proposed geometry interpolation combines random geometry portions from generated and real pattern distributions to form gradient-penalty samples.
- Customized interpolation: The interpolated samples fully characterize patterns between the generated and real distributions.
- Customized interpolation: The interpolation method allows the generator to extrapolate and explore the ground-truth distribution when training data do not cover the whole design space.Training experiments evaluate training stability, design accuracy, and extrapolation capability.
4. Hyperparameters and training curves
The GAN training curves stabilize as the models learn to generate samples close enough to real data, while intermediate models progressively produce more refined meta-atom shapes and focused fields.
- Training curves: After approximately 3,000 epochs, generator and discriminator losses stabilized for the GANs.At this point, the discriminator could not differentiate generated samples from real samples.
- Training curves: Training curves for the meta-atom design network converged after 10,000 iterations.
- Training evolution: The study recorded intermediate GAN models after 1, 2, 100, and 3,000 training iterations to visualize design evolution.These models were used to design and numerically test the same bifocal metalens.
- Training evolution: During training, generated meta-atoms evolved from large, unclear shapes into more diverse patterns with refined details.
- Training evolution: The corresponding E-field distributions gradually converged to two sharp focal spots as training progressed.The focal spots emerged in the design based on the 3,000-iteration model.
5. Polarization-independent metalens design
The dual-polarization generative network designs a polarization-independent transmissive focusing lens by specifying phase profiles for both polarizations. Full-wave simulations show equal 80 μm focal lengths with near-equal field magnitudes.
- Design approach: A polarization-insensitive transmissive lens can use an equal focal length of 80 μm for both polarizations.
- Design approach: Equal x- and y-polarization phase shifts primarily produce structures with 4-fold rotational symmetry.The symmetry requirement can be relaxed when the relative phase difference remains constant.
- Design approach: A 90 degree phase bias was added to the x-polarized phase mask while designing the lens.
- Design approach: The dual-polarization network generates one qualified meta-atom for each lens cell from target phase profiles under both polarization incidences.
- Results: Full-wave simulations produced focal fields with the same 80 μm focal length and near-equal magnitude for both polarizations.