Learning to inversely design acoustic metamaterials for enhanced performance

Abstract

Elastic metamaterials are popularly sought to realize numerous special functions such as vibration control and wave manipulation among which sound absorption is a typical task fulfilled by acoustic metamaterials. Inverse designing metamaterials with machine learning approaches has been under the spotlight thanks to the data-driven experience-free advantages and become one of the important design paradigms. Nevertheless, the existing works mostly concentrate on validating the reproduction accuracy of the neural networks on trained data and very few have explored their ability on designing for enhanced properties. To this end, our work studies the competence of the proposed inverse design framework in enhancing the acoustic performance of a three-dimensional mixed-size cavity-based waterborne sound absorptive metamaterial. With forward and inverse networks in the framework, the target sound absorption spectra (100-10000 Hz) are taken as inputs into the inverse network during training and a corresponding structure is output with the best matching spectra which is subsequently fed into the forward network for acoustic property evaluation and loss calculation. The trained forward network is shown to possess excellent generalization capabilities by highly accurately predicting for structures with “unseen” beyond-range parameters compared to the training set. Most importantly, the inverse network is delightfully capable of spontaneously adopting beyond-range structural parameters to ensure meeting the acoustic target whose mean sound absorption coefficient is higher than any of the data in the training set, hence inverse designing for enhanced performance. The inverse design accuracy is dramatically improved from only 9.2% of mean squared errors being <0.0001 to 99.6% with beyond-range exploration. A case study is presented to demonstrate the significant difference beyond-range exploration makes for inverse designing aiming at enhanced performance. It is hoped that this work will serve as an inspiration for the design and optimization of elastic metamaterials with enhanced performance for future work.

摘要

弹性超材料可被用于实现多种特殊功能, 例如振动控制和波操纵, 其中吸声是声学超材料完成的典型任务. 利用机器学习方法反向设计超材料因利用数据驱动且不依赖经验的优势而备受关注, 成为重要的设计范式之一. 但现有的工作大多集中在验证反向设计神经网络的设计准确性, 很少有人探索如何利用神经网络进行反向设计以提高材料性能. 为此, 我们的工作研究了反向设计框架在提升3D多尺度空腔型声学超材料的声学性能方面的能力. 框架中有正向和反向网络, 在训练过程中将目标吸声曲线(100~10000 Hz)作为反向网络的输入, 并输出具有满足此吸声曲线的相应超材料结构, 随后将其输入正向网络进行声学性能评估. 结果表明, 经过训练的前向网络可对超过训练集结构参数范围的结构(既“未见过”的结构)进行高精度性能预测, 因此具有较高泛化性能. 更重要的是, 反向网络能够自发地采用超出范围限制的结构参数, 以确保满足平均吸声系数高于训练集中任何数据的吸声目标, 因此可利用反向设计突破训练集中数据的声学性能, 进行性能优化. 通过超限探索, 反向设计精度从仅有9.2% 的设计拥有<0.0001的均方误差显着提高到99.6%. 最后, 利用案例证明神经网络拥有的超限探索能力对旨在提高性能的反向设计有重要意义. 希望这项工作能为更高性能弹性超材料的优化设计提供支撑.