Optomechanical reservoir computing

Abstract

Nonlinear dynamics are pervasive phenomena in natural and synthetic material systems, where time-varying signals from different physical stimuli in the environment influence the material system behavior. Physical reservoir computing leverages these nonlinear dynamics to produce complex input–output mappings by interpreting the dynamical system as a physical recurrent neural network. A source of physical nonlinearity is crucial for enabling the reservoir to predict nonlinear relationships. Despite the significance of nonlinearity, most physical reservoirs leverage only a single source of nonlinearity. Furthermore, there exists a gap between analyses that examine fundamental capabilities of reservoir computers and those that evaluate the practical performance of reservoir computers. In this study, an optomechanical reservoir is introduced that combines both the nonlinear dynamics from bilinear springs and nonlinear sensing from optical fibers. Both the nonlinear springs and the optical fibers are shown to contribute significantly to the range of nonlinear frequency content produced by the optomechanical reservoir. A novelty search of simulated reservoirs highlights the range of performance exhibited by the optomechanical reservoir, and several high performing designs are validated experimentally. Additionally, a frequency content metric is introduced to characterize the nature of a given reservoir’s nonlinearity, highlighting what kinds of frequencies the reservoir can and cannot produce. This analysis is an important step toward the rational design of reservoir computers as it allows one to match reservoir computers with computational tasks. The development of both analytical techniques and multiphysics designs lays the groundwork for more effective embodied intelligence in dynamic systems.