OPTIMIZATION OF PLATE PHONONIC CRYSTALS WITH MULTI-SHAPE HOLES VIA DEEP NEURAL NETWORKS

Designing phononic crystals with wide bandgaps has been the focus of extensive researches. However, the designed shapes based on conventional methods are relatively simple, and mainly focusing on single-shape unit cells. This study optimizes plate phononic crystals with multi-shape holes to further exploit bandgap broadening potentialities, and the optimization is achieved through deep neural networks (DNN). The initial dataset consists of a huge number of randomly generated unit cells with multi-shape holes. Simulation models are built using the commercial software COMSOL to calculate the bandgaps of the unit cells. For each unit cell, a DNN is trained to predict its bandgap, while another DNN is used to classify the presence or absence of a bandgap. Utilizing the predictive capabilities of the DNN, a genetic optimization algorithm is employed to search for shapes with broad bandgap, which are then validated through COMSOL simulations. The DNN is then updated using active learning techniques based on these analyses results. The final results demonstrate that the DNN-based design of plate phononic crystals with multi-shape holes is effective, and they exhibits wider bandgaps compared to those with single-shape holes.