Deep learning-based inverse design of irregular phononic crystals

Abstract

Designing phononic crystals (PCs) with irregular scatterer geometries is a computationally intensive and challenging task, typically requiring iterative optimization and extensive numerical simulations. Here, we propose a deep learning (DL) framework capable of predicting geometric structures of PCs from specified dispersion properties. Our model comprises a convolutional autoencoder (AE) and a fully connected deep neural network (DNN). The AE extracts low-dimensional features from scatterer images and reconstructs predicted geometric structures in the inverse design process, while the DNN establishes the complex mapping between these geometric features and their corresponding band structures. Once trained, this model enabling the dispersion prediction and inverse design of the scatterers thus can be used to efficiently design PCs with desired properties, such as bandgaps and Dirac cone dispersions. Based on the DL model, we demonstrate the inverse design of a zero-refractive-index PC exhibiting a dispersion of triple Dirac point dispersion. Our study offers a robust and efficient tool for the design of PCs with complex scatterers, opening up new avenues for applications in acoustics, materials science, and beyond.