Combining Material and Geometric Nonlinearity for Optomechanical Reservoir Computing

Nonlinear dynamics are ubiquitous across engineering systems. Physical reservoir computing leverages these dynamics to perform useful computations within the physics of the application. The physical nonlinearity in the dynamic system (i.e., reservoir) enriches the information from incoming signals, allowing the reservoir to function as a recurrent neural network when linear regression is performed on the high dimensional reservoir state. Often, a single source of nonlinearity is leveraged. However, we recently demonstrated that combining nonlinear springs in a 1D mass-spring-damper system with the nonlinear transmission through bending optical fiber readouts provided a greater range of nonlinear computational ability. In this study, we add a third nonlinearity to the system. We study 2D mass-spring-damper networks in simulation where the dynamic angle changes in the reservoir produce geometric nonlinearity. We leverage an ablation study and our previously developed frequency content analysis to characterize the effect of combining these nonlinearities on the reservoir's performance. Additionally, we show that smaller node degrees and moderate variation in the dynamic spring angular orientation tend to produce greater nonlinearity. The insights from this study will likely prove useful guidelines for future reservoir design.