Learning to Compute with Sound in Nonlinear Disordered Cavities

Machine learning (ML) is now widely used for a variety of important tasks. To improve ML performance, several computational architectures have been investigated whose physical implementations promise compactness, physical robustness, and low energy cost. Here, we experimentally demonstrate an approach that uses a nonlinear disordered cavity as a complex reverberant acoustic environment for the physical realization and enhancement of the computational power of a type of ML known as reservoir computing (RC). We realize the required fading memory of RC systems by leveraging the internal memory of sound-propagation and the intrinsically large feedback of disordered structures to increase the expressive power and reduce the total size of proposed RC. In addition, the use of reconfigurable nonlinear electroacoustic resonators on the cavity's top side allows us to enhance the reservoir size and, as a result, the performance of the acoustical reservoir computer. We show the efficiency of our approach on a challenging benchmark problem: forecasting chaotic time series.