Ultrathin acoustic absorbing metasurface based on deep learning approach

Acoustic metasurface has become one of the most promising platforms for manipulating acoustic waves with the advantage of ultra-thin geometry. The conventional design method of acoustic metasurface relies on numerical, trial-and-error methods to retrieve effective properties of the locally resonant unit cells. It is often inefficient and requires significant efforts to investigate the enormous number of possible structures with different physical and geometric parameters, which demands huge computational resources. This is especially when modeling narrow cavities where thermoviscous loss has to be considered. In this paper, a deep learning-based acoustic metasurface absorber modeling approach is introduced to significantly reduce the characterization time while maintaining accuracy. Based on a convolutional neural network (CNN), the proposed network can model wide absorption spectrum response in the timescale of milliseconds. The performance of the implemented network is compared with other classical machine learning methods. Using CNN, we have demonstrated an ultrathin metasurface absorber having perfect absorption at an extremely low frequency of 38.6 Hz with an ultrathin thickness down to λ/684 (1.3 cm). The total path length for the propagating waves inside the channel is about λ/5.7 which breaks the quarter-wavelength resonator theory. The network prediction is validated using the experiments to demonstrate the effectiveness of this physical mechanism. Furthermore, we propose a broadband low-frequency metasurface absorber by coupling unit cells exhibiting different properties based on the supercell concept. This approach is attractive for applications necessitating fast on-demand design and optimization of a metasurface acoustic absorber. (paper)