A deep learning-based multi-fidelity optimization method for the design of acoustic metasurface

A desirable acoustic metasurface requires the scattered acoustic field distribution uniform. Neural networks are effective substitutions to mimic the expensive FE simulations in most research. However, the computational cost required to construct a model with only single high-fidelity (HF) simulation data is still unacceptable. This paper presents a deep learning-based multi-fidelity optimization framework to improve the uniformity of the scattered acoustic field distribution. First, a multi-fidelity composite convolutional neural network (MF-CCNN) method is proposed to predict the high-dimensional scattered acoustic field at a lower data cost. The developed MF-CCNN consists of four convolutional subnets. The first part predicts a low-fidelity (LF) output, whose features are then extracted by the second part and concatenated with the inputs to predict the HF result. Two parallel branches are utilized to map the LF features to the HF output. Then, the physical parameters' optimization neural network is proposed to minimize the objective under the prediction of MF-CCNN. The proposed method is compared with other state-of-the-art multi-fidelity networks, and the results demonstrate that MF-CCNN reaches the highest accuracy and the mean absolute error is improved by at least 20%. The variance of the obtained scattered acoustic field after optimization is reduced by 3.62%, and the time cost is only 8% of the genetic algorithm (GA), proving the efficiency and accuracy of the proposed framework.