Prediction of acoustic transmission and kinetic energy partition in chiral metamaterials using deep neural networks
Abstract
Acoustic metamaterials (AMMs) can be designed for the control of sound wave scattering and propagation for specific frequencies and angles of incidence to achieve a desired performance. Chiral AMMs, whose geometry breaks mirror symmetry, are of interest to control acoustic transmission for their ability to convert energy between translational and rotational motion. This work first considers the effects of geometry and material properties of a chiral AMM by modeling acoustic reflection and transmission through an AMM layer using finite element analysis (FEA). FEA enables analysis of the energy distribution within the AMM unit cell to determine the relative magnitudes of translational and rotational kinetic energy components to elucidate the role of translational and rotational energy partition in acoustic transmission. However, performing FEA for all possible variations in design parameters can incur prohibitively high computational costs. To address this bottleneck, the present study develops fast surrogate models to enable efficient design space exploration for chiral AMMs. Specifically, deep neural networks, which can be trained on limited FEA results, are utilized to predict transmission and kinetic energy components as functions of frequency, incidence angle, and design parameters. Necessary constraints on FEA training data are discussed, as well as associated computational expenses.