Absorption characteristics prediction and inverse design of acoustic metamaterial via deep learning

Conventional design strategies for acoustic metamaterials necessitate extensive expertise and specialized acumen on the part of designers, entailing significant computational expenditures in the design process. The merits of deep learning methodologies, leveraging a data-centric paradigm without dependence on empirical designs, have attracted escalating attention in the contemporary academic discourse. Consequently, this investigation delves into its application for the design of acoustic metamaterials. We employed a Convolutional Neural Network (CNN) to address the forward prediction challenge, attaining an impressive prediction accuracy of up to 98%. Building upon these results, we formulated a conditional Deep Convolutional Generative Adversarial Network (c-DCGAN) to confront the inverse design conundrum. The network is capable of swiftly generating topological structures that align with predetermined absorption curves, thereby demonstrating robust generalization prowess in inverse design scenarios. It is proficient in executing inverse design tasks accurately for a diverse array of previously unseen types of feature curves.