Inverse design of two-dimensional tunable acoustic valley-Hall insulators via deep learning

Abstract

Acoustic valley-Hall insulators leverage topological protection properties to enable unidirectional waveguiding, defect immunity, and low-loss transmission, exhibiting significant potential in applications such as energy harvesting, sensing, and signal processing. In the present work, a novel deep learning model is developed for designing topological configurations of acoustic valley-Hall insulators to simultaneously satisfy dual requirements of target Dirac point frequencies and temperature parameters. The deep learning model consists of a generation module and an evaluation module. The generation module uses a variational autoencoder to generate candidate topological configurations. The evaluation module employs a four-level physical constraint processing system, which implements four-stage progressive screening through the frequency domain screening layer, physical field verification layer, feature compression layer, and regression prediction layer. The two modules form a closed-loop optimization through dynamic feedback. The results show that the model can inversely generate corresponding topological configurations of acoustic valley-Hall insulators for given Dirac point frequencies and temperature parameters with high accuracy. Compared to conventional trial-and-error approaches constrained by the combinatorial complexity of parameters, this method enhances design efficiency while enabling exploration across high-dimensional design spaces. Finally, based on the unit cell of acoustic valley-Hall insulators obtained from the inverse design model, a topological valley-locked waveguide of sound waves was realized.