Accelerated design of low-frequency broadband sound absorber with deep learning approach

Broadband sound absorption has been a long-lasting concern in the field of noise engineering, yet to date remains challenging to cover a broad low-frequency range with ultra-thin materials of a few centimeters. A practicable approach is using coherently coupled resonators to constitute a compact coplanar metasurface absorber. However, this scheme leads to a tough inverse problem posed by the large number of design parameters since all have to be meticulously tuned to satisfy the critical coupling condition. We tackle this problem with the deep learning approach. An autoencoder-like neural network is built that, once maturely trained, significantly promote the inverse design process thanks to the highly efficient data-driven-based forward and inverse predictions. In the design, we have added a probabilistic model into the neural network to enhance its robustness for the normally ill-posed inverse design problems which require artificial and probably unreal spectrum as an input. This probabilistic network is capable of providing multiple ultra-thin (32 mm) and broadband metasurface designs. The optimized designs have been numerically and experimentally verified, showing the capacity of using solely nine resonators to achieve quasi-perfect sound absorption (absorption coefficient α⩾0.9) in a band from 350 to 530 Hz. Our work is helpful to accelerate the design of metasurface absorbers targeted especially for broadband noise control at low frequencies.