An efficient inverse design framework of manufacturable phononic metaplates via combining generative deep learning and Bayesian latent space exploration

Abstract

While deep learning (DL) techniques emerge as powerful tools for inverse designs of phononic metamaterials, prevailing DL-driven approaches heavily rely on predicting structure–property relationship that is very time-consuming in data collection and model training. Moreover, most reported methodologies have only been conceptually validated in numerical models of 2D planar phononic crystals (PnCs), with limited implementation in practical, manufacturable metastructures. In this study, we present a prediction-free framework to efficiently design manufacturable phononic metaplates via a combination of generative DL model and Bayesian space exploration strategy. The generative DL model is used to construct the design space and encode unit cell patterns into low-dimensional latent vectors. Thereafter, Bayesian optimization is employed to efficiently search for desirable structures in the encoded latent space, where unit cells satisfying user-customized dispersion properties can be discovered after a small set of iterations. Band gap properties of inversely designed metaplates are demonstrated by the experimentally measured wave transmission. Our developed framework greatly enhances the design efficiency of manufacturable metaplates and is generalizable to 2D metamaterials.