Inverse design of locally resonant metabarrier by deep learning with a rule-based topology dataset

Abstract

A deep learning model consisting of a variational autoencoder (VAE) and a tandem neural network (TNN) is presented to realize the inverse design of a locally resonant metabarrier for vibration mitigation in low-frequency region. The optimized topological configuration and periodic constant are taken as the design variables, and a soil parameter is treated as a condition variable considering its influence on bandgaps. To improve the quality of samples and control the design space reasonably, Configuration Generation (CG) rules are formulated to guide the generation of a topology dataset. The design performance of the deep learning methodology is tested by a large number of samples based on numerical simulations, and it is found that the designed bandgaps are highly consistent with the targeted ones. Multiple designed results are given for the same target, revealing the “one-to-many” nature of the design problem. Two metabarriers are designed for vibration mitigation in the main frequency ranges of a practical example, and considerable blocking effects are observed.