Sound absorption of resonant composite metastructure based on machine learning reverse-assisted design

• We proposed a resonant composite metastructure for reverse-assisted design. • Use the first 10,000 peaks generated by varying the metastructure's geometry as the dataset. • Two-stage and one-stage DNN were adopted to predict the geometric parameters. • Two-stage DNN achieves a high-precision of absorption coefficients and frequencies. To realize machine learning reverse-assisted design for resonant sound-absorbing structures, this study introduces resonant composite metastructures, which are constructed from perforated sheets, cavities, insert plates, and porous materials. The sound absorption coefficients within the range of 10,000 Hz were theoretically derived, and the underlying sound absorption mechanisms were thoroughly discussed. A dataset comprising 100,000 randomly generated first peaks was constructed using Latin hypercube sampling to support machine learning applications. Two approaches, a one-stage deep neural network and a two-stage deep neural network incorporating a forward prediction component—were developed to inversely predict the structural dimension parameters of eight resonant sound-absorbing units based on 24 sets of desired first peak characteristics. The results demonstrated that the two-stage model significantly outperformed the one-stage approach, achieving markedly higher accuracy in predicting both the frequency and sound absorption coefficient of the first peaks. The effectiveness of the machine learning predictions was further validated through acoustic impedance tube experiments on two samples designed via the two-stage deep neural network. These findings underscore the potential of machine learning for the efficient and accurate reverse design of resonant sound-absorbing structures.