Master of Science in Electrical Engineering, University of Dayton, 2023, Electrical and Computer Engineering

Memristors are among the leading devices for non-volatile memory applications due to its in-memory computing capability, power efficiency and high endurance, as well as the speed at which memory states can be written. Tantalum oxide is currently one of the most promising materials in memristor design. Tantalum oxide based memristor devices operate via oxygen vacancy migration in the TaOx layer forming a filament which then allows current flow once formed. Due to the migration of the vacancies through the material to form this chain, there is permanent structural damage to the device over the course of operation which eventually results in reduced performance of the devices and less change between the SET/RESET states. In this paper, we analyze the impact of zirconium doping concentrations on oxygen vacancy migration by studying the electric field required to induce switching in the devices and device endurance. We test the power efficiency and ON/OFF resistance ratios of numerous TiN/TaOx/Ta/TiN devices made with zirconium oxide doping layers. The results are then analyzed to understand the effects of the TaOx/ZrO2 bilayer structure and the impact on the total power efficiency and endurance of the devices. The devices did show increased conductivity but switching performance and endurance were reduced. TEM imaging was performed to determine the cause, but no definitive cause could be identified.

Committee: Guru Subramanyam (Committee Chair); Andrew Sarangan (Committee Member); Ganguli Sabyasachi (Committee Member) Subjects: Electrical Engineering

MS, University of Cincinnati, 2020, Engineering and Applied Science: Electrical Engineering

The characterization of HfO2 ReRam devices is critical in understanding their switching properties and potential applications. From the known switching properties, an organization of observed phenomena allows for development of a concise framework of device operation. The developed framework acts as a pivot point for understanding and effectively modeling the switching properties in a concise manner. Once this framework is developed, a physical insight into the action of the device can be truly obtained. The method of modeling ReRam devices follows this approach. The physics of device action is understood from the testing data and matched by the framework, allowing for the development of the model. After the model has been fully developed, the testing data trends are shown to match the developed model and are explained to be naturally accommodated by the developed framework. During this process, useful properties of ReRam devices came to light. These properties are shown to have a potential application in hardware security.

Committee: Rashmi Jha Ph.D. (Committee Chair); Marc Cahay Ph.D. (Committee Member); Luke Duncan PhD (Committee Member) Subjects: Electrical Engineering

MS, University of Cincinnati, 2018, Engineering and Applied Science: Electrical Engineering

This thesis work expanded on traditional resistive random access memory (ReRAM) technology, which relies on movement of oxygen ions through a metal oxide, with a gate terminal to better control ion motion. This terminal allows for modulation of the ion gradient in the metal-oxide region, controlling the device conductivity. The advantages of such a device over traditional ReRAM lie in the independence of read and write operations, lower power consumption, and higher state range. The practical implications of these advantages over traditional ReRAM are better control over device writing and improved flexibility for certain applications, as well as lower energy consumption and additional available discrete states, allowing for higher density. The advantages of ReRAM in general over traditional non-volatile memory (NVM) technologies such as flash are high endurance, low voltage requirements, and fast write speeds. The gate-controlled device was designed based on commonly accepted conduction mechanisms in metal oxides, which was further developed into a full microfabrication process flow. Photomasks were designed to explore the device characteristics at different physical dimensions. Fabrication was done using RF magnetron sputtering for all depositions, photolithography for patterning, and wet etchants and reactive ion etching (RIE) for etching. The fabricated device was thoroughly characterized for set/reset characteristics, endurance, DC and AC behavior, and parasitics. These device measurements were used to develop a behavioral device model which is based on both experimental results and physics-based mechanisms, and was fit to the experimental data on several devices to obtain a large distribution of device characteristics. Finally, this model was used to explore the usage of such a device as a non-volatile memory cell and in analog artificial neural networks for embedded machine learning. The fabricated device demonstrated the desired behavior, capable of independ (open full item for complete abstract)

Committee: Rashmi Jha Ph.D. (Committee Chair); Marc Cahay Ph.D. (Committee Member); Cory Merkel Ph.D. (Committee Member) Subjects: Computer Engineering

MS, University of Cincinnati, 2018, Engineering and Applied Science: Computer Engineering

Software and hardware artificial neural networks have been increasingly used to solve difficult real world problems. Neural networks have shown successful results in both the High Performance Computing and Big Data fields, where computation is often distributed among clustered hardware, such as CPUs and general purpose Graphic Processing Units (GPUs), to form much larger computing systems. The growing field of neuromorphic computing combines aspects of complex biological systems and classical machine learning methods, accelerated by alternative physical platforms. The field exemplifies a complementary class of neural network computing platforms that aim to emulate neural and synaptic dynamics performed by the human brain with solid-state devices designed to mimic the electro-chemical behavior of their counterpart. These systems can be exceedingly compact and low-power compared to current machine learning methods performed on distributed CPUs and GPUs, and solve similar problems. The physical scalability of neuromorphic platforms is promising due to the advent of material and fabrication technology of both volatile and non-volatile memory, such as resistive RAM (ReRAM). In this thesis, a framework for simulating spiking neurons, ReRAM devices used as synapses, and neural network models is demonstrated and analyzed for biologically-inspired neuromorphic computing research and applications. Additionally, a software framework is designed to handle and simulate various dynamics of ReRAM models used as synapses in complex neural network architectures. A collection of software and mathematical modules are defined to model neuromorphic devices and circuit-level dynamics, as well as, provide a pipeline for analyzing both simulated and collected device data for use in training and testing complex neuromorphic circuitry. Finally, an neuromorphic architecture is proposed and evaluated as dense networks of spiking neuron groups connected by synapses to learn from supervised d (open full item for complete abstract)

Committee: Rashmi Jha Ph.D. (Committee Chair); Marc Cahay Ph.D. (Committee Member); Ali Minai Ph.D. (Committee Member) Subjects: Electrical Engineering