Emergent Physical Intelligence in Biomimetic Scale Metabeams

Abstract

Physical reservoir computing (PRC) leverages the intrinsic dynamics of physical systems to perform information processing while requiring training only in a linear readout layer. Here, we introduce a geometry-programmable metabeam inspired by the hierarchical overlapping architecture of biological scales as a physical reservoir. Contact-mediated interactions between embedded scales induce tunable nonlinear dynamics, enabling controlled transitions between periodic, multi-periodic, and chaotic responses under external excitation. The computational capability of the metabeam reservoir is evaluated using static nonlinear function approximation, Lorenz-63 prediction, and the NARMA-2, NARMA-5, and NARMA-10 benchmarks. Input signals are encoded through the excitation amplitude, whereas computation is realized through the vibrational amplitude of the metabeam followed by a trained linear readout. Across all benchmark tasks, the proposed metabeam consistently outperforms an equivalent linear beam reservoir, yielding lower normalized root-mean-square errors (NRMSEs) and improved computational performance. Different dynamical regimes of the metabeam exhibit distinct computational advantages, with the periodic regime providing the highest prediction accuracy for memory-intensive tasks and the multi-periodic regime achieving the best performance in static nonlinear function approximation. These findings demonstrate that simple interacting surface textures can simultaneously enrich reservoir dynamics and information processing capability, establishing scale-covered metabeams as tunable and mechanically programmable platforms for embodied physical intelligence and PRC.