Probabilistic deep neural network-based inverse design method for the underwater acoustic metasurface of customizable absorption performance

Abstract

Underwater acoustic metasurfaces have attracted significant research attention for their subwavelength thickness and superior acoustic wave manipulation performance, for example, broadband sound absorption down to the low frequency range. However, at present, efficient and automatic inverse design of underwater absorption acoustic metasurfaces is still a challenging task due to the unidentifiability and highly nonlinear relationship between spectral responses and structural parameters. This brings a barrier to the practical applications of metasurface-based acoustic devices. In this work, an inverse design method for the underwater acoustic metasurface of the on-demand absorption spectrum is investigated via a probabilistic deep neural network. By introducing a probabilistic generation module, the proposed neural network improves the generalization capability of the inverse design and allows for the generation of diverse candidate structure parameters for the Helmholtz resonator array with embedded apertures. The effectiveness and accuracy of the proposed approach are validated by our theoretical analysis and simulations. This strategy enriches the diversity and flexibility of the design solutions and possesses the merit of good tolerance to fabrication errors, which effectively facilitate the inverse design of underwater acoustic metasurface aiming at sound absorption and noise control.