A CuOx/p+-Si memristor with short- and long-term plasticity for homogeneous reservoir computing system

Abstract

Reservoir computing (RC) system is a simple and cost-efficient neuromorphic computing system mainly consisting of a reservoir layer and a readout layer. The reservoir layer comprises nonlinear nodes with short-term plasticity (STP), while the readout layer comprises linear nodes with long-term plasticity (LTP). Here, we propose a self-rectifying (>10²) memristor based on CuO_x/p⁺-Si heterojunction that exhibits both nonlinear STP and linear LTP characteristics with high uniformity (σ/μ = 2.8 %). The resistive switching of the device is mainly based on electron trapping/detrapping in CuO_x film. The trapping before relaxation is very unstable and less capable of long-term trapping, which leads to STP. The trapping after relaxation is relatively stable for the lowered trap barrier, which induces LTP. Utilizing its dynamic STP characteristics for the reservoir layer and LTP properties for the readout layer, a CuO_x/p⁺-Si memristor-based homogeneous dynamic RC system was constructed for a spoken-digital recognition task, yielding an accuracy rate of 95.33 %. These results affirm the tremendous potential of our device in establishing a highly compact and full-memristor-based homogeneous RC system.