Bifunctional Memristor with Neuronal-Synaptic Coupling for Passively Self-Adaptive Reservoir Computing

Physical reservoir computing (RC) is highly efficient for temporal signal processing. However, the robustness of physical RC systems can be limited by the fixed dynamics of physical nodes, which may enter saturation under wide-range inputs. To address this issue, we propose an Ag/Ti/TaO_x/Pt memristor that functionally couples neuronal (dynamic response) and synaptic (memory tunable) behaviors under different electrical operations. When two such devices are connected in an anti-series configuration, they form a reservoir node that passively adapts to wide-range inputs without the need for active regulation circuits. Specifically, as the input signal approaches saturation, the synapse's voltage division increases, triggering a SET transition that reduces its voltage and consequently raises the neuron's voltage, thus expanding the effective dynamic range. We evaluated this passive-adaptive reservoir on a chaotic Hénon Map prediction task. The results demonstrate a nearly 80% reduction in average normalized root-mean-square error (NRMSE) compared to a conventional dynamic memristor reservoir across 110 different input conditions. This work provides a promising neuron-synapse coupled hardware-efficient node for building robust physical RC systems, with potential applications in complex temporal signal processing and edge computing.