Nanoscale HfO2-based memristive devices for neuromorphic computing

Abstract

Redox-type memristive devices (ReRAM) based on ultrathin metal/metal oxide/metal stacks are considered as one of the most promising approaches for future high-density non-volatile data storage and beyond-von Neumann architectures, including emerging fields of storage class memory, machine learning and neuromorphic computing (NC) [1]. Fast switching events on time scales of nanoseconds combined with data retention times in the order of ten years are enabled by the nanoscale redox-type reactions in the ReRAM cells that control the addressable resistance states. Stack design and device operation in metal oxide-based valence change mechanism (VCM)-type memristive devices is understood from the perspective of oxygen transfer as well as drift/diffusion processes [2]. The realization of highly dense packed arrays seems feasible due to the inherent potential of the devices for scalability and three-dimensional integration making ReRAM cells interesting as artificial synapses in NC circuits. In the context of neuromorphic computing the accessibility of intermediate resistance states is one of the important requirements [3]. In addition, resistive switching devices with volatile resistance states are of increasing interest for implementation of NC learning rules [4]. Energy-efficient memristive devices for in-memory-computing (IMC) and next generation NC applications must fulfill requirements like compatibility with CMOS back-end-of-line (BEOL) and three-dimensional fabrication, device scalability, and operation parameters fitting to the design node of the circuitry. Hafnium oxide is one of the most promising materials for CMOS-compatible ReRAM devices. In VCM-type memristive cells, the HfO2 film sandwiched between an inert and a reactive metal electrode acts as the actively switching layer, where the resistance change occurs due to oxygen ion movement between the conductive HfOx filament and the insulating HfO2–x disc region.