Mechanical Computing with Magneto-Mechanical Oscillators

The rapid advancement of artificial intelligence is stimulating the exploration of design strategies to embody intelligent processes into dynamic, nonlinear material and structural responses. Physical reservoir computing is a promising approach that leverages the nonlinear dynamics of physical systems to perform computational tasks, offering advantages in speed, energy efficiency or insertion opportunities over traditional digital computing. In this study, we investigate the information processing utility of the nonlinear interactions of magneto-mechanical oscillators as a physical reservoir. Our model system consists of a 1D array of repulsive magnets that are each connected to a base plate via a cantilever beam. As the base plate is vibrated, the inertial response, finite cantilever deflections and the repulsive magnetic interactions synergize together to produce a rich state matrix of magnet positions vs time. The design space of spring and magnetic properties is computationally surveyed using a novelty search optimization to discover configurations with robust frequency content. We then apply linear regression on these magnet trajectories to generate target output computations, such as benchmark tasks of sine delay, nonlinear moving averages, and input signal classification. Together, these results demonstrate the types of nonlinear mappings possible with magneto-mechanical interactions and their utility for embodied signal processing.