Harnessing Nonlinearity for Optomechanical Reservoir Computing

Physical reservoir computing is a computing paradigm that performs machine learning via a nonlinear, dynamical system. Reservoir computers (RCs) have leveraged a variety of physics, harnessing the unique nonlinearites. The nonlinearity is critical to the RC's design as it increases the dimension of the input signal. However, connection between the physical nonlinearity and computing performance is not yet well defined. We develop a spectral projection method to characterize this relationship between system nonlinearity and computing performance, and evaluate it in an optomechanical RC. This analysis partitions the dimensionality increase of the signal among its frequencies, distinguishing between the strengths of the linear and various classes of nonlinear responses. The nonlinear springs in the RC produce bilinear force-displacement responses. Elastomeric optical fibers coupled to the springs bend with spring compression, nonlinearly changing the optical transmission due to Snell's law. The RC's performance is shown both in simulation and experiment. This work 1) demonstrates the benefit of combining different physical nonlinearities and 2) develops greater insight into the connection between the RC nonlinearity and computing performance, leading to more efficient RCs.