Controlling Ionic Transport in RRAM for Memory and Neuromorphic Computing Applications

Abstract

Resistive random-access memory, based on a simple two-terminal device structure, has attracted tremendous interest recently for applications ranging from non-volatile data storage to neuromorphic computing. Resistive switching (RS) effects in RRAM devices originate from internal, microscopic ionic migration and the associated electrochemical processes which modify the materials’ chemical composition and subsequently their electrical and other physical properties. Therefore, controlling the internal ionic transport and redox reaction processes, ideally at the atomic scale, is necessary to optimize the device performance for practical applications with large-size arrays. In this thesis we present our efforts in understanding and controlling the ionic processes in RRAM devices. This thesis presents a comprehensive study on the fundamental understanding on physical mechanism of the ionic processes and the optimization of materials and device structures to achieve desirable device performance based on theoretical calculations and experimental engineering. First, I investigate the electronic structure of Ta2O5 polymorphs, a resistive switching material, and the formation and interaction of oxygen vacancies in amorphous Ta2O5, an important mobile defect responsible for the resistive switching process, using first-principles calculations. Based on the understanding of the fundamental properties of the switching material and the defect, we perform detailed theoretical and experimental analyses that reveal the dynamic vacancy charge transition processes, further helping the design and optimization of the oxide-based RRAM devices. Next, we develop a novel structure including engineered nanoporous graphene to control the internal ionic transport and redox reaction processes at the atomic level, leading to improved device performance. We demonstrate that the RS characteristics can be systematically tuned by inserting a graphene layer with engineered nanopores at a vacancy-exchange interface. The amount of vacancies injected in the switching layer and the size of the conducting filaments can be effectively controlled by the graphene layer working as an atomically-thin ion-blocking material in which ionic transports/reactions are allowed only through the engineered nanosized openings. Lastly, better incremental switching characteristics with improved linearity are obtained through optimization of the switching material density. These improvements allow us to build RRAM crossbar networks for data clustering analysis through unsupervised, online learning in both neuromorphic applications and arithmetic applications in which accurate vector-matrix multiplications are required. We expect the optimization approaches and the optimized devices can be used in other machine learning and arithmetic computing systems, and broaden the range of problems RRAM based network can solve.