Related Experiment Video
Updated: May 11, 2026
09:33
Demonstration of Equal-Intensity Beam Generation by Dielectric Metasurfaces
Published on: June 7, 2019
6.3K
A surrogate-assisted extended generative adversarial network for parameter optimization in free-form metasurface
Manna Dai1, Yang Jiang2, Feng Yang1
1Computing and Intelligence Department, Institute of High Performance Computing, Agency for Science, Technology and Research (A*STAR), 138632, Singapore.
Summary
This study introduces XGAN, a deep learning model that rapidly designs complex free-form metasurfaces for 5G communication. XGAN accelerates metasurface design by 500x with high accuracy, enabling efficient library creation.
Area of Science:
- Electromagnetics and Metamaterials
- Artificial Intelligence in Engineering
Background:
- Metasurfaces are crucial for 5G microwave communication.
- Free-form metasurfaces offer superior spectral response control over regular shapes.
- Traditional design methods for free-form metasurfaces are slow and require expertise.
Purpose of the Study:
- To develop an accelerated and accurate method for designing free-form metasurfaces.
- To leverage deep learning for inverse design of metasurfaces.
- To create a physically constrained generative adversarial network (GAN) for metasurface synthesis.
Main Methods:
- An extended generative adversarial network (XGAN) was developed.
- A surrogate model with physical constraints was integrated into XGAN.
- XGAN was trained to generate metasurfaces directly from spectral responses.
Main Results:
- XGAN achieved an average accuracy of 0.9734 in generating 20,000 free-form metasurface designs.
- The proposed method is 500 times faster than conventional numerical techniques.
- The XGAN approach ensures accurate monolithic generation of metasurfaces.
Conclusions:
- XGAN significantly accelerates the design of high-quality free-form metasurfaces.
- This method facilitates the construction of metasurface libraries for specific spectral requirements.
- The approach is extensible to diverse inverse design challenges in optics and materials science.
