Deep-learning-based inverse design of phononic crystals for anticipated wave attenuation

Bandgaps of phononic crystals dominating the propagation of evanescent waves have received significant attention recently, which can be determined and tuned by the topology of a unit cell. Predicting a band structure and designing topological structures with desirable characteristics have become a research hotspot. In this study, a data-driven deep learning framework is applied to arrive at the prediction of the band structure and the inverse design of topology. A convolutional neural network is trained to predict band structures of phononic crystals. After training a generative adversarial network, the generator is concatenated with the convolutional neural network for inverse design. Meanwhile, a complex band structure of phononic crystals is computed by the periodic spectral finite element method to present the spatial decay of evanescent waves. The topology with the greater spatial attenuation is screened from the ground truth topology and the inversely designed topology. Finally, an optimized topological phononic crystal with an anticipated bandgap is obtained, which has the potential for better acoustic insulation and vibration isolation.