Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bidirectional Adversarial Autoencoders for the design of Plasmonic Metasurfaces

Yuansan Liu, Jeygopi Panisilvam, Peter Dower, Sejeong Kim, James Bailey

TL;DR

This paper addresses the ill-posed inverse design problem for plasmonic metasurfaces, where a one-dimensional spectrum must map to a two-dimensional geometry. It introduces the Bidirectional Adversarial Autoencoder (BiAAE), a two-stream latent architecture that jointly models spectrum and metasurface shapes via common and unique latent components and four loss terms, enabling implicit spectrum–shape mappings without explicit conditioning. The approach yields nonlinear, multi-peak designs and improves generative quality (lower FID by ~30% versus cDCGAN) while maintaining reconstruction capability, as validated by Lumerical FDTD simulations. The work advances nonlinear metasurface design and offers a scalable framework for other dimensionality-mismatch inverse design problems, with potential enhancements through richer data representations and additional optical constraints.

Abstract

Deep Learning has been a critical part of designing inverse design methods that are computationally efficient and accurate. An example of this is the design of photonic metasurfaces by using their photoluminescent spectrum as the input data to predict their topology. One fundamental challenge of these systems is their ability to represent nonlinear relationships between sets of data that have different dimensionalities. Existing design methods often implement a conditional Generative Adversarial Network in order to solve this problem, but in many cases the solution is unable to generate structures that provide multiple peaks when validated. It is demonstrated that in response to the target spectrum, the Bidirectional Adversarial Autoencoder is able to generate structures that provide multiple peaks on several occasions. As a result the proposed model represents an important advance towards the generation of nonlinear photonic metasurfaces that can be used in advanced metasurface design.

Bidirectional Adversarial Autoencoders for the design of Plasmonic Metasurfaces

TL;DR

This paper addresses the ill-posed inverse design problem for plasmonic metasurfaces, where a one-dimensional spectrum must map to a two-dimensional geometry. It introduces the Bidirectional Adversarial Autoencoder (BiAAE), a two-stream latent architecture that jointly models spectrum and metasurface shapes via common and unique latent components and four loss terms, enabling implicit spectrum–shape mappings without explicit conditioning. The approach yields nonlinear, multi-peak designs and improves generative quality (lower FID by ~30% versus cDCGAN) while maintaining reconstruction capability, as validated by Lumerical FDTD simulations. The work advances nonlinear metasurface design and offers a scalable framework for other dimensionality-mismatch inverse design problems, with potential enhancements through richer data representations and additional optical constraints.

Abstract

Deep Learning has been a critical part of designing inverse design methods that are computationally efficient and accurate. An example of this is the design of photonic metasurfaces by using their photoluminescent spectrum as the input data to predict their topology. One fundamental challenge of these systems is their ability to represent nonlinear relationships between sets of data that have different dimensionalities. Existing design methods often implement a conditional Generative Adversarial Network in order to solve this problem, but in many cases the solution is unable to generate structures that provide multiple peaks when validated. It is demonstrated that in response to the target spectrum, the Bidirectional Adversarial Autoencoder is able to generate structures that provide multiple peaks on several occasions. As a result the proposed model represents an important advance towards the generation of nonlinear photonic metasurfaces that can be used in advanced metasurface design.
Paper Structure (9 sections, 7 equations, 5 figures)

This paper contains 9 sections, 7 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Data Generation
  4. Inverse Design via GANs
  5. Implicit Relationship Learning
  6. Results
  7. Qualitative Results
  8. Quantitative Results
  9. Conclusion

Figures (5)

  • Figure 1: A set of randomly generated structures using the cDCGAN architecture
  • Figure 2: Overview of Bidirectional Adversarial Autoencoder
  • Figure 3: Results for the implementation of our system. the red dashed line indicates the predicted spectrum for our corresponding meta-atom shape, where the black line indicates the input spectrum on which the neural network generated a structure. (a) The result shows a structure with a nonlinear response characteristic as seen by the multi peak output in response to a single peak input (b) An example of a nonlinear response characteristic from a generated structure with sharper resonant peaks (c) An example of a multi peak resonance generation (d) A single peak generation based on an input spectrum (e) Shape reconstruction using the BiAAE
  • Figure 4: Mean Squared Error ($\downarrow$) with varying sample size.
  • Figure 5: Kernel Density Estimation of FID Score ($\downarrow$)