Committee Members:
Dr. Eric Vogel, Advisor, MSE
Dr. Meilin Liu, MSE
Dr. Matthew McDowell, MSE
Dr. William Alan Doolittle,  ECE
Dr. Samuel Graham, University of Maryland, ME/Dean of Engineering

Impact of thermal material properties and local ion concentration on Ti/HfOx-based analog devices

Abstract: As we near the scaling limit for silicon-based transistors, developing new materials and devices for non-traditional computing becomes increasingly important. Today’s computer architecture suffers from the von Neumann bottleneck, in which computational instructions share the same pathway as data writes. This architecture limits system efficiency, speed, and scaling. Analog adaptive oxide devices, commonly referred to as memristors or resistive random-access memories (RRAMs), avoid this bottleneck and have garnered significant attention for machine learning and neuromorphic compute in-memory circuits. The realization of adaptive oxides for neuromorphic computing hinges on repeatable and predicable analog changes of electrical resistance, which is fundamentally controlled by the materials in and around the device. This work focuses on filamentary memristors which exhibit a non-volatile change in resistance by modulating the concentration of oxygen vacancies within a small (filamentary) region of an otherwise insulating oxide layer in a metal-insulator-metal (MIM) stack. Under an applied electric field, these devices experience localized temperature rises over 1000 K on picosecond timescales, with drift, diffusion, and thermophoresis causing the migration of oxygen ions and oxygen vacancies. All three of these mechanisms have a strong dependence on temperature. Therefore, the management of the thermal field is crucial to successful implementation of these materials and devices. This dissertation independently establishes the impact of the substrate and electrode thermal conductivity both experimentally and computationally. For biologically realistic pulse widths, low substrate thermal conductivities led to increased resistance changes in RRAM devices. Furthermore, scanning thermal microscopy was used to compare the in-situ temperature rise of the top electrode directly above the filament with the estimated value from the model. This established a method to estimate the filament temperature during biasing with an accuracy ~30 K.  Computational results demonstrated the temperature of the capping layer (between the oxide and the top electrode) had the greatest impact on the resistance change. Thus, a low thermal conductivity capping layer led to significantly higher resistance changes. Further work exploring the importance of the capping layer revealed that slightly higher initial oxygen concentrations (~2 - 3%) caused larger resistance changes compared to lower concentrations. In summary, this work establishes the importance of the thermal properties not only in contact with the filament, but also far away (substrate and electrodes) and establishes the importance of understanding the interplay between the filament and the capping layer to further improve the analog resistance change of filamentary RRAMs.