The 2D-global optimization networks (2D-GLOnets) model for acoustic metamaterial design

Publication Date


Document Type

Conference Proceeding

Publication Title

Proceedings of the International Congress on Acoustics


This study aims to contribute to the development of innovative data-driven approaches for metamaterial design using global optimization, deep learning, and generative modeling. The 2D-Global Optimization Networks (2D-GLOnets) technique is implemented to design acoustic metamaterials. The 2D-GLOnets optimize the probability of providing globally optimized meta-devices in the design space. A reparametrization technique is developed and is applied to constrain the scatterers into a feasible region. The gradients of the loss (objective) function with respect to the weights are calculated from backpropagation, and they update the weights of the generator. The physics-based gradients are calculated analytically from a multiple scattering solver [1] and used in backpropagation. The 2D-GLOnets generate a batch of metamaterials in each iteration and calculate the loss (objective) function over the entire training. The 2D-GLOnets gradually use the information provided to the model. The method is employed for sound localization and acoustic lens design. In a broadband acoustic lens design, the 2D-GLOnets maximize the absolute pressure at the focal point at discrete wavenumber values. The method is illustrated with examples of planar configurations of cylindrical scatterers producing sound localization and focusing effects.


Acoustic Metamaterials, Generative Modeling, Global Optimization, Machine Learning


Mechanical Engineering