基于机器学习的声子晶体结构声隐身设计

Vibration optimization design of sonar hood based on machine learning

摘要:

针对声呐导流罩中低频段振动控制以提高其全频段声隐身性能问题，提出了一种基于条件自编码器的声子晶体结构逆向设计方法，在目标频段逆向设计出的具有带隙的声子晶体结构可作为夹芯结构芯层，为声呐导流罩声振特性治理工作提供新思路。首先随机生成大量声子晶体周期单元，并提出两种策略扩充在目标频段内具有带隙的样本数量。针对有限元软件批量计算声子晶体结构带隙效率较低的问题，训练了卷积神经网络用于识别声子晶体是否具有带隙。最后将声子晶体结构以及带隙分布作为训练集训练条件自编码器，结果显示，卷积神经网络对结构的带隙具有很好的识别效果，识别准确率可以达到89%；条件自编码器能学习到人工周期结构的轴对称结构，生成的人工周期结构与原结构仅相差几个像素，且生成结构的带隙与原结构带隙误差小于1%，说明本文方法可以应用于声子晶体结构的逆向设计。

Abstract:

A reverse design of phonon crystal structure based on conditional autoencoder is proposed in response to the low-frequency vibration control issue in the sonar hoods. The designed phonon crystal structure, featuring a bandgap in the target frequency band, can be used as the sandwich structure's core layer, providing a new idea for the vibroacoustic characteristics management of the sonar dome. Firstly, many phonon crystal periodic units are randomly generated, and two strategies are proposed to expand the number of samples with band gaps in the target frequency band. To solve the problem of low efficiency of batch calculation of phonon crystal structure bandgap by Finite Element Software, a convolutional neural network is trained to identify whether phonon crystal has bandgap. Finally, the phonon crystal structure and bandgap distribution are used as the training condition autoencoder. The results show that the convolutional neural network has a good recognition effect on the band gap of the structure, and the recognition accuracy can reach 89%. The conditional autoencoder can learn the axisymmetric structure of the artificial periodic structure. The generated artificial periodic structure is only a few pixels different from the original structure. The band gap error between the generated structure and the original structure is less than 1%, indicating that this paper's method can be applied to the reverse design of the phonon crystal structure.