Inverse design of laminated plate-type acoustic metamaterials for sound insulation based on deep learning

Abstract

The laminated plate-type acoustic metamaterial (LPAM) has designable low-frequency sound insulation performance and ultrathin and ultralight characteristics. However, the complexity of LPAM design increases significantly as the number of layers increases. An inverse design method is proposed to determine the topology and design parameters of LPAM based on deep learning. To train the deep learning model, the finite element and analytic models are combined to generate sufficient training data. The experimental validation of the finite element model and the acoustic impedance model are carried out to reduce the deviation between the experiments and simulations. The LPAM design system contains the pre-processing, inverse design and post-processing modules. The pre-processing module can produce candidate targets based on the required sound-insulation design target. For each candidate target, the inverse design module utilizes convolutional neural networks to autonomously design topology and parameters of LPAM. Finally, the pre-processing and post-processing modules are combined to choose the optional LPAMs meeting the sound-insulation target. To verify effectiveness of the LPAM design system, typical cases, involving the sound insulation targets of one and two sound insulation peaks, and uniform sound insulation in a broadband from 50 Hz to 600 Hz, are designed successfully and efficiently.