Prediction and inverse design of compact hybrid acoustic metamaterial absorbers enabled by deep learning

Abstract To address the urgent demand for low-frequency, broadband, and compact noise control structures in modern engineering, this paper proposes an intelligent design framework based on deep learning for hybrid acoustic absorbers combining micro-perforated panels and space-coiling structures. This hybrid design achieves excellent impedance matching at deep subwavelength scales by folding the acoustic propagation paths within a compact structure. To overcome the high computational costs associated with conventional finite element method simulations in multi-parameter optimization, a high-fidelity dataset was constructed via Latin hypercube sampling. Based on this, a forward prediction network (decoder) and an inverse design network (encoder) utilizing a Tandem Network architecture were developed. The results demonstrate that the forward network achieves a coefficient of determination ( R 2 ) of 0.9939 on the test set, enabling the prediction of full sound absorption spectra in milliseconds. The inverse network successfully addresses the ‘non-uniqueness’ challenge in acoustic design, precisely deducing structural parameters from user-defined absorption targets. Finally, experimental samples were fabricated using 3D printing and laser cutting technologies. Impedance tube tests show high agreement with simulation results, validating the significant potential of this intelligent framework for delivering customized noise control solutions.