Modeling acoustic metamaterials with manufacturing uncertainties using deep learning

Approximate models and the inevitable imperfections introduced by additive manufacturing (AM) continue to pose significant challenges to the reliability of acoustic metamaterial (AMM) performance. While these limitations are well acknowledged within the research community, focused efforts to directly mitigate their impact remain very limited. This paper presents a deep learning framework to bridge the gap between what is designed, what is expected, and what is actually obtained. A structured latent autoencoder (SLAE) is proposed and trained on a dataset obtained from the theoretical model of a meta-absorber. Due to observed mismatches between theoretical and experimental results, a data augmentation technique is introduced to incorporate random transformations of the original spectra within informed ranges including scaling, shifting, noise addition, and bandwidth modulation. This approach offers a two-fold benefit: compensating for imperfections in the theoretical model due to unknown physics/approximation errors, and simulating manufacturing impacts on the acoustic response. Subsequently, the augmented SLAE is integrated into a transfer learning (TL) framework to account for real-world manufacturing artifacts. The proposed method not only closes the gap between idealized AMMs and their real-world counterparts, but also features a reliable and cost-effective AMM solutions.