Current analog-tunable memristors often suffer from poor retention, limiting their role as nonvolatile synaptic nodes in hardware-based neuromorphic networks. On the other hand, conventional nonvolatile memristors display abrupt switching, necessitating additional current-regulating devices for circuit function. To address these challenges, we developed vertically arranged Bi2Se3 memristors suitable for crossbar array configuration via gold-assisted, plasma-etching-free physical vapor deposition (PVD), facilitating scalable production. These memristors exhibit notable analog conductance tuning (10-40%) with minimal postsetting relaxation and stable nonvolatile retention (4 s) without the need for a current regulator (e.g., a current-limiting transistor). They operate at low switching voltages (-0.98 V SET, +0.7 V RESET). The superior switching attributes stem from the electromigration-driven filament formation in the Au/Bi2Se3/Ti structure. Utilizing these memristors, we built an all-hardware-based reservoir computing network where Bi2Se3 memristors serve effectively in the readout layer, processing sensory signals and generating control commands with minimal digital electronics. This configuration achieves ultralow power consumption (∼7 μW), significantly lower than digital equivalents, while maintaining competitive performance, evidenced by a normalized root-mean-square error (NRMSE) of 0.094. These results position Bi2Se3 memristors as robust, energy-efficient options for next-generation neuromorphic computing.
Ki et al. (Sun,) studied this question.