Machine Learning Driven Inverse Design of Broadband Acoustic Superscattering

Acoustic superscattering holds great potential in applications such as acoustic imaging, nondestructive testing, and sound wave sensing. Recently, machine learning attracted increasing attention in the study of acoustic metamaterials. Here, we apply machine learning techniques to the design of acoustic superscatterers with multilayer core–shell structures, aiming to enhance broadband sound scattering performance. Deterministic and probabilistic deep learning models based on autoencoder architectures are developed to retrieve the structural and material parameters of superscatterers. A forward neural network is first constructed to predict the scattering cross‐section spectrum from the design parameters, while a loss function of weighted mean absolute error is introduced to improve the prediction accuracy at spectral peaks and valleys. Subsequently, an inverse neural network is established to realize the mapping from scattering spectra to structural information. The probabilistic model improves the generalization and robustness of the design process and facilitates practical fabrication. This approach provides an efficient framework to accelerate the inverse design of intelligent acoustic superscattering devices with customizable broadband responses and offers a feasible solution to the inverse scattering problem. Broadband high‐insulation from the normalized frequency of 0.85–1.4 is achieved with the designed superscatterer.