Deep reinforcement learning for optimal sound absorbing structures design

Abstract

Deep learning algorithms have demonstrated a tremendous success in designing structures that surpass human capabilities. Based on the recent achievements of deep reinforcement learning in surpassing human capabilities, this paper focuses on implementing these algorithms to design optimal configurations of solid and porous materials that achieve a broadband absorption within the frequency range of 300 Hz to 3000 Hz. We employ model-free approaches, specifically deep Q-learning, double deep Q-learning, and dueling deep Q-learning algorithms, to predict material configurations that optimize absorption without requiring expertise knowledge. From a 230×30 different material combinations, the deep reinforcement algorithms learn to predict configurations that yield optimal absorption in few hundred steps. We discuss further the superior performance of a dueling deep learning algorithm compared to the other two deep learning approaches and a heuristic approach, such as genetic algorithm. The proposed model-free algorithms enable the prediction of absorption performance for any material configurations without the need for expertise.