Accelerated engineering of topological interface states in one-dimensional phononic crystals via deep learning

TL;DR

This paper presents a deep learning framework for the rapid and flexible design of one-dimensional phononic crystals with specific topological interface states, enabling instant inverse design and robustness verification.

Contribution

It introduces a combined autoencoder and neural network approach for high-dimensional PnC design, achieving over 97% prediction accuracy and enabling one-to-many TIS frequency design.

Findings

Successful inverse design of PnCs with specific band gaps and TIS frequencies.

High prediction accuracy (>97%) for PnC properties.

Instantaneous and reliable PnC design verified experimentally.

Abstract

Topological interface states (TISs) in phononic crystals (PnCs) are robust acoustic modes against external perturbations, which are of great significance in scientific and engineering communities. However, designing a pair of PnCs with specified band gaps (BGs) and TIS frequency remains a challenging problem. In this work, deep learning (DL) approaches are used for the engineering of one-dimensional (1D) PnCs with high design freedoms. The considered 1D PnCs are composed of periodic solid scatterers embedded in the air background, whose unit cell is divided into a matrix with 32 * 32 pixels. First, the variational autoencoder is applied to reduce the dimensionality of unit cell images, allowing accurate reconstruction of PnC images with different numbers of scatterers. Subsequently, the multilayer perceptron and the tandem neural network are used to realize the property prediction and customized design of 1D PnCs, respectively. The correlation coefficients for the property prediction and inverse design are more than 97%. The unit cell images of 1D PnCs with specific BG properties could be successfully and instantaneously designed. Importantly, the implementation of a "one-to-many" design of PnC pairs with specific TIS frequencies is realized. Furthermore, the reliability and robustness of the constructed networks are confirmed by randomly specifying the design targets as well as the experimental verification. This study demonstrates the broad application prospects of DL approaches in the field of PnC design and provides new ideas and methods for the intelligent design of artificially functional materials.