Non-local generative machine learning-based inverse design for scattering properties.

Area of Science:

• Electromagnetism
• Materials Science
• Computer Science

Background:

• Metamaterials manipulate electromagnetic waves using arrays of scatterers.
• Current design methods for metasurfaces are limited in geometric structures, materials, and electric field control.
• Arbitrary electric field generation is a key challenge in metamaterial design.

Purpose of the Study:

• To propose an inverse design method for metamaterials using generative adversarial networks (GANs).
• To overcome limitations of traditional metamaterial design by enabling arbitrary electric field distributions.
• To improve the efficiency and quality of electric field generation in metamaterials.

Main Methods:

• Developed a GAN-based inverse design framework with a forward model and inverse algorithm.
• The forward model uses dyadic Green's function to map scattering properties to electric fields.
• The inverse algorithm employs computer vision techniques to transform scattering properties and electric fields into images for GAN training.

Main Results:

• The proposed GAN architecture with ResBlock effectively designs metamaterials for target electric field patterns.
• Achieved significantly greater time efficiency compared to traditional metamaterial design methods.
• Generated higher quality electric fields, demonstrating optimal scattering properties for specific field patterns.

Conclusions:

• The GAN-based inverse design method offers a powerful tool for creating advanced metamaterials.
• This approach overcomes limitations in geometric and material diversity for metasurface design.
• The method validates the potential for generating arbitrary and high-quality electric fields using optimized scattering properties.