A new inverse design method for sound-absorbing metamaterial based on deep learning

Abstract

Recent advances in deep learning demonstrate significant potential for accelerating the design of complex acoustic absorbers. However, current approaches predominantly utilize complete absorption spectra as network inputs and rely on fixed unit cell configurations during design phases. These constraints introduce inefficiencies and practical limitations in inverse design implementations. To address these challenges, we present an innovative deep learning framework for inverse design applications, and apply it to the design of a subwavelength acoustic metamaterial with multiple micro-slit resonators. Distinct from conventional methods, our approach requires only target sound absorption indices (defined by lower and upper frequency bounds) as neural network inputs. Furthermore, the system enables adaptive adjustment of the number of unit cells according to prescribed absorption bandwidth requirements, significantly enhancing design flexibility and practicality. For efficient dataset generation, we establish a theoretical model revised via Kriging surrogate technique. Comparative analyses of deep neural networks (DNN) and convolutional neural network (CNN) reveal that both architectures achieve accurate predictions of metamaterial structural parameters across the 370–1200 Hz frequency range. Experimental validations confirm the effectiveness of our developed strategies, while subsequent discussions address the generalization abilities of neural networks. This investigation represents a substantive progression in deep learning-driven inverse design strategies for acoustic metamaterial absorbers.