Descriptors to Dynamics: A Materials and Device Perspective on in-Materio Physical Reservoir Computing for Neuromorphic Edge Intelligence.

Abstract

Reservoir computing is attractive for edge intelligence because it not only reduces training overhead but also turns transient material physics into computation. In in-materio systems, memory, nonlinearity, separability, and efficiency are not added later at the circuit level; they emerge from defect kinetics, ionic motion, interfacial barriers, percolation pathways, and related dynamics. This perspective therefore treats material descriptors as design variables rather than post hoc explanations of device behavior. Accordingly, six design rules connect measurable material properties to useful reservoir function: timescale matching, operating-regime nonlinearity, controlled stochasticity, observable state dimensionality, reversible and energy-relevant operation, and robustness across an operating range. To make these rules actionable, the framework introduces a staged experimental workflow, a minimum quantitative characterization set, attribution controls separating material dynamics from input encoding and readout design, and a cross-layer reporting checklist. The focus remains material- and device-based in-materio physical reservoirs for edge and in-sensor processing, while the descriptor-to-dynamics logic is transferable to other physical reservoirs. The framework is then extended to reservoir-by-sensing platforms, where sensing and temporal processing coexist within one substrate. Finally, reporting of operating conditions, state observables, stability, and energy boundaries supports fair comparison and helps move proof-of-concept demonstrations toward deployable materials systems.