Accelerated inverse design of ultra-thin broadband sound absorber based on data-driven method

This study introduces a data-driven inverse design method for hybrid microperforated panel-homogeneous porous material (MPP-HPM) composite structures to achieve broadband sound absorption. Initially, a deterministic autoencoder-based model is utilized, within which the decoder is trained to map geometric parameters to sound absorption spectra, while the encoder is trained to deduce the geometric parameters from given spectra. The deterministic model demonstrates high-precision bidirectional mapping with a low mean absolute error. Nevertheless, due to the uncertainties caused by artificial or non-physical input spectra, a probabilistic model is incorporated to introduce a latent space to overcome the limitations of deterministic models. The key findings indicate that the probabilistic model is capable of generating multiple feasible solutions for any user-defined input of the inverse problem. Moreover, the closer the user-defined input is to the initial spectra, the more efficiently inverse solutions can be obtained. The proposed approach achieves high-efficiency broadband absorption with average α = 0.932 within the target absorption band (1035–5135 Hz) while featuring an ultra-thin thickness of merely 37 mm. This research paves the way for more efficient noise control strategies in various engineering and industrial environments.