SUMMARY

Memristors are considered as one of the promising candidates for next-generation computation and storage systems. Neuromorphic computing architectures based on memristors were demonstrated to be very efficient in neural network training. Memristors made of transition-metal-oxides, like HfOx and TaOx, achieve resistive switching by formation and rupture of conductive filaments. The formation of filaments is driven by diffusion, drift, and electrophoresis of oxygen vacancies, which are affected by electric field, Joule heating, etc. When voltage bias is applied in the presence of conductive filament, high current and resulting temperature rise may result in structural change driven by thermal gradients and electric field. The physics of memristor is not completely understood, and there are many challenges for their employment in commercial applications. Computational modeling of memristor operation requires accurate material properties such as thermal conductivity and diffusion coefficient, many of which are not available. A better understanding of the material properties and high-fidelity modeling techniques will push this field to a new stage. This research first investigates the structural and diffusion properties of amorphous memristor material at different temperatures through molecular dynamics simulations. Two recently developed potential parameter sets - MBKS and COMB - were used for the simulations to investigate properties of amorphous hafnium dioxide and also to examine their predictive capabilities. Next, a frequency domain thermoreflectance (FDTR) set-up is developed to measure the thermal properties of materials used in memristor devices. A deep learning based data analysis pipeline is built to improve the measurement confidence by avoiding uncertainty due to convergence to local minima during curve fitting. Using the FDTR system, the thermal conductivity and volumetric heat capacity of HfOx thin films with various thicknesses were investigated. Positive temperature dependence of thermal conductivity was observed in HfOx thin films. The difference in volumetric heat capacity among measured samples is probably due to different structures resulting from different deposition conditions. In-operando thermal mapping with sub-micrometer spatial and sub-microsecond temporal resolutions was used on functioning tantalum oxide memristive switches, and hot spots were observed corresponding to oxygen concentration gradients, indicating the presence of localized conductive filaments. A hybrid electro-thermal model comprising of 3-D heat transfer and 0-D resistive switching model was constructed and calibrated against the measurements to predict electrical characteristics and temperature rise. Thermal crosstalk in memristor crossbar was demonstrated using the hybrid model. Such a model will guide system design considering thermal performances, which is critical to most future electronic chips.