Conditional Wasserstein generative adversarial networks applied to acoustic metamaterial design

Area of Science:

• Computational physics
• Materials science
• Applied mathematics

Background:

• Minimizing total scattering cross section (TSCS) is crucial in wave scattering problems.
• Traditional methods like forward modeling are computationally intensive.
• Efficient methods are needed for optimizing scattering properties.

Purpose of the Study:

• To present a novel deep learning-based method for reducing the TSCS of planar cylinder configurations.
• To leverage generative modeling for efficient TSCS minimization.
• To demonstrate the efficacy of the proposed method for 2D scattering problems.

Main Methods:

• Utilized generative modeling and deep learning, specifically conditional Wasserstein generative adversarial networks (cWGANs).
• Integrated convolutional neural networks (CNNs) with cWGANs to simulate TSCS.
• Enhanced the cWGAN model with a coordinate convolution (CoordConv) layer for improved performance.

Main Results:

• The cWGAN model successfully generates 2D cylinder configurations that minimize TSCS.
• The deep learning approach offers a more computationally efficient alternative to repeated forward modeling.
• Demonstrated the method's effectiveness using examples of planar uniform cylinder configurations.

Conclusions:

• Generative modeling combined with deep learning provides an efficient solution for TSCS reduction.
• The proposed cWGAN model with CoordConv layer is effective for optimizing scattering properties.
• This work opens new avenues for designing materials with tailored wave interaction characteristics.