Inverse design of membrane-type acoustic metamaterial based on deep learning

Abstract

Membrane acoustic metamaterial (MAM) design mainly relies on forward design methods, making it difficult to determine structural parameters directly from predefined acoustic requirements. In this study, an inverse design framework for rubber-metal composite mass-block MAM (RMMAM) is proposed. A LiveLink-based COMSOL–MATLAB simulation interface is established to automatically generate finite element datasets, and a dual attention convolutional neural network based on the EfficientNet-feature pyramid architecture is trained to predict structural parameters from target sound transmission loss (STL) curves. Unlike conventional forward strategies, the proposed framework directly maps target STL curves to the corresponding RMMAM parameters. With target STL curves as input, the prediction error of structural parameters is controlled within 10.5%. In addition, STL curves reconstructed from the predicted parameters exhibit peak-frequency errors of 5–20 Hz and STL errors of 1.13–10.92 dB, demonstrating the reliability of the inverse design model. Case studies are further conducted under the constraint that the peak sound insulation level should not be lower than 50 dB in the 300–400 Hz frequency band. Both numerical simulations and experiments confirm that the designed RMMAM structures satisfy the predefined sound insulation requirements. The proposed framework provides an efficient route for the inverse design of acoustic metamaterials and supports their engineering applications.