Machine learning-assisted low-frequency and broadband sound absorber with coherently coupled weak resonances

An artificial broadband sound absorber composed of multiple components is of significant interest in the physics and engineering communities. The existence of coherently coupled weak resonances (CCWRs) makes it difficult to achieve optimal broadband sound absorption, especially in the presence of complex and aperiodic components. Here, we present and experimentally implement a machine learning-assisted subwavelength sound absorber with CCWRs using an improved Gauss–Bayesian model, which exhibits flexible, high-efficient, and broadband properties at low frequencies (<500 Hz). The proposed aperiodic structure comprises three parallel split-ring units, which enable a quasi-symmetric resonant mode to be generated and effectively dissipate energy because of the huge phase difference between each component at the coupled resonant frequency. With high algorithmic efficiency (no more than 80 iterations), the improved Gauss–Bayesian model inversely determines the optimal CCWRs, realizing a reconfigurable high sound absorption spectrum (α > 0.9) from 229 to 457 Hz. The optimal configuration of sound spectrum characteristics and the unit cell structure can be confirmed flexibly. Good agreement between numerical and experimental results verifies the effectiveness of the proposed method. To further exhibit broadband and multiparameter optimization, a nine-unit sound absorber (27 parameters) is numerically simulated and shown to achieve high acoustic absorption and a relatively broad bandwidth (44.8%). Our work lifts the restrictions on analytic models of complex and aperiodic components with coherent coupling effects, paving the way for combining machine learning with the optimal design of metamaterials.