Vibroacoustic Metamaterial Design via Deep Reinforcement Learning

Abstract

Optimization or trial-and-error approaches are commonly employed to design vibroacoustic metamaterials to effectively isolate vibration in the low and mid frequencies. Particularly, gradient-based and gradient-free optimizations result in a single optimal structure that meets the specified constraints. However, multiple designs satisfy these constraints and solutions within the given frequency ranges. In this paper, we demonstrate that employing a deep reinforcement learning approach allows for generating various designs that fulfill the specified constraints, i.e., designing a lightweight vibroacoustic metamaterial while maximizing the band gap. The computational efficiency, geometric generation, and stability of the algorithm are discussed in detail. The flexibility of our proposed approach marks a significant advancement in automating the design of vibroacoustic metamaterials, taking into account mass and manufacturing constraints for real-world applications.