Characteristics and Applications of Non-Volatile Resistive Switching (Memristor) Device

Non-volatile memory technology scaling has been driven by the ever increasing needs of high-capacity and low-cost data storage. Scaling the conventional floating gate device structure, however, has faced with several technical challenges due to constraints of electrostatics and reliability. Alternative memory approaches based on non-transistor structures has been extensively studied. Among the new approaches, resistive switching devices (RRAM) have attracted tremendous attention due to their high endurance, sub-nanosecond switching, long retention, scalability, low power consumption, high ON/OFF ratio and CMOS compatibility.

In this thesis, we present a systematic study on the fundamental understanding and potential applications of RRAMs. Firstly, we introduce a quantitative and accurate model of the dynamic resistive switching processes, by solving the coupled equations for oxygen vacancy transport, current continuity and Joule heating. Secondly, we show systematic investigations on the resistance switching mechanism through detailed noise and transport analysis, and develop a unified model to explain the conduction path and account for the resistance switching effects. Thirdly, we perform detailed retention studies of oxide-based RRAMs at elevated temperatures and develop an oxygen diffusion reliability model of RRAM devices. The activation energy for oxygen vacancy diffusion is directly calculated from the measurement. Analytical modeling and detailed numerical multi-physics simulation is discussed. Fourthly, we report that doping tantalum oxide based RRAM with silicon atoms leads to larger dynamic ranges with improved accessibility to the intermediate states which is suited for neuromorphic computing applications. Lastly, we investigate the application of RRAMs in neuromorphic computing by showing data clustering based on unsupervised learning. Through both simulation and experimental studies, we demonstrate that a crossbar array of RRAM devices can perform data clustering through unsupervised learning and enable effective data classification in a real-world problem.

These studies have not only helped the development and optimization of RRAM devices but also highlighted their application potential beyond simple memory. We believe continued development of this emerging device structure may lead to future high-performance and energy efficient memory and logic hardware systems.