Reservoir Computing with Active Matter Systems

Reservoir computing with physical systems offers a promising route toward next-generation and in materio computing. Lymburn et al. (Chaos 31(3), 033121, 2021) recently introduced reservoir computing based on simulations of active matter. However, the optimal properties of such systems remain unclear. Here, we systematically investigate the effects of (1) external input injection, (2) intrinsic agent dynamics and control, (3) agent-agent interactions, and (4) coarse-grained readout design. For a Lorenz-63 chaotic time series prediction task, we find that optimal performance occurs around a critical damping threshold. Predictive accuracy correlates with velocity correlations and their fluctuations, with the near-critically damped regime remaining robust across different chaotic attractors and parameter variations. Remarkably, these optimal dynamics already emerge in single-particle reservoirs. Active matter forming liquid droplets with sharp interfaces around the input proves particularly effective. These insights advance the design of soft-matter reservoirs and unconventional computing devices.