Dual-Mode Memristors with Controlled Cluster Dynamics for High-Performance Neuromorphic Computing

Abstract

In the process of realizing brain-like information processing based on artificial synapses, the conductive filament (CF) model has become an intensively studied topic. Its significant attraction stems from its ability to emulate the dynamic evolution of synaptic connection strengths in biological systems. However, owing to its inherent geometrical structural defects, traditional CF exhibits significant uncontrollability during its growth and dissolution processes, severely limiting the stability and accuracy of its synaptic function simulation. To address this limitation, a synaptic device was constructed in this study that features a dual-active-layer composed of polyvinylpyrrolidone (PVP): graphene-oxide quantum dots (GO QDs), and polyvinyl alcohol (PVA). The thin PVA active layer enables compliance-current-controlled regulation of dendritic conductive filaments, resulting in volatile switching at low compliance currents and nonvolatile switching at higher compliance currents, thereby enabling compliance-current-controlled dual-mode operation. Ag ions originating from the electrochemically active Ag electrode migrate into the active layer driven by the external electric field and are reduced in situ around the GO QDs in the PVP:GO QDs nanocomposite layer, forming uniformly distributed and controllable silver clusters (Ag-C), resulting in enhanced device stability and improved synaptic emulation capability. The proposed mechanism is consistent with the voltage-dependent SEM/EDS observations and previously reported charge-trapping characteristics of GO QD-containing materials. Experimental results confirmed that diverse synaptic behaviors can be realized in the device. Meanwhile, a neural-network simulation based on its synaptic characteristics demonstrated its excellent performance (94.0%) in learning and inference tasks. Such performance further highlights the potential of the device for constructing reliable neuromorphic computing architectures.