Accelerated inverse design of customizable acoustic metaporous structures using a CNN-GA-based hybrid optimization framework

Abstract

A key challenge in sound-absorbing structure design is the effectively inverse generation of structure capable of achieving high absorption at desired frequencies. As the two of the most commonly used methods in inverse design, trial-and-error and traditional topology optimization suffer extensive computational costs. In this work, we develop a CNN-GA hybrid optimization framework that combines the advantages of the traditional heuristic algorithm with emerging machine learning technology to accelerate the inverse design of a random sound-absorbing structure called metaporous with only 30mm thickness based on the sound absorption in diffuse field. The finite element method (FEM) simulation is applied to calculate the sound absorption coefficients of randomly generated metaporous structures. The numerous data obtained from the FEM simulation are input to the convolutional neural network (CNN) model for training the network. The well-trained CNN model, of which generalization ability is validated by a testing set, is coupled with a genetic algorithm (GA) for accelerating the iterative evaluation of the objective function in inverse design. The proposed inverse design framework leverages the accelerated GA to generate near-optimal metaporous structures of various tailored absorption peaks typically in roughly 5 s to 30 s on average. The numerical results show that the proposed CNN-GA hybrid optimization framework works several orders of magnitude faster than the conventional single algorithms in inverse design of metaporous, on the condition that accuracy is ensured. In addition, an impedance tube measurement is performed to validate the proposed method. This work brings forward a new inverse design method for metaporous materials and brings out the immense potential of combining the meta-heuristic algorithm and machine learning.