Exploiting Bistability and Viscoelasticity in Reservoir Computing

Abstract Reservoir computing is a powerful framework for processing time-domain signals, leveraging the nonlinear dynamics of physical systems to solve complex tasks. A key requirement for a nonlinear dynamical system to function effectively as a reservoir computer (RC) is its ability to retain recent inputs while gradually forgetting older ones. In electrical systems, this short-term memory is easily implemented using a delay line. However, in mechanical systems, viscoelasticity is the only mechanism for information retention, as the system’s state depends not only on its current deformation but also on its deformation history. Motivated by this, we investigate the role of viscoelasticity in mechanical reservoir computers and explore how tuning its parameters can enhance prediction performance. In this paper a chain of nonlinear oscillators composed of viscoelastic bistable unit cells is used as the reservoir computer and highlights the important memory-nonlinearity trade-off in an RC. The nonlinearity is introduced through multistability and the short term memory through viscoelasticity. Benchmark tasks including the N-bit parity operation and the NARMA-10 task are performed to show the efficacy of these kinds of oscillator chains for computation. The memory capacity of the system is quantified, and the influence of parameter variations on predictive accuracy is systematically analyzed. We further demonstrate that when the reservoir exhibits chaotic dynamics, long-term predictions deteriorate. To address this, we apply the Melnikov criterion to approximately identify parameter regimes for good RC performance. Finally, the reservoir computing functionality of the bistable chain is demonstrated through a scaled down analog electrical implementation.