An Efficient Acoustic Metamaterial Design Approach Integrating Attention Mechanisms and Autoencoder Networks

Abstract

Acoustic metamaterials have been widely applied in fields such as sound insulation and noise reduction due to their controllable band structures and unique abilities to manipulate low-frequency sound waves. However, there exists a highly nonlinear mapping relationship between their structural parameters and performance responses, which causes traditional design methods to face the problems of inefficiency and poor generalization. Therefore, this paper proposes a bidirectional modeling framework based on deep learning. We constructed a forward prediction network that integrates an attention mechanism, a multi-scale feature fusion, and a reverse design model that combines an improved autoencoder and cascaded neural network to efficiently model the dispersion performance of acoustic metamaterials. In the feedforward network, the improved forward prediction model shows superior performance compared to the traditional Convolutional Neural Network model and the model based only on the Convolutional Block Attention Module attention mechanism, with a prediction accuracy of 99.65%. It has better fitting ability and stability in the high-frequency part of the dispersion curve. In the inverse network part, compression of the high-dimensional dispersion curves by an improved autoencoder reduces the training time by about 13.5% without significant degradation of the inverse prediction accuracy. The proposed network model provides a more efficient method for the design of metamaterials.