Since the end of Dennard Scaling, the scaling of a silicon transistor's power with its size, the semiconductor industry has been investigating alternative kinds of computer architectures in order to continue its historic gains. Neuromorphic architectures powered by networks of reconfigurable resistors are one such system. The need to develop tunable, analog resistive memories has helped drive a resurgence of interest in oxide resistive switching, a poorly understood phenomenon studied on and off over the decades and widely believed to be driven by defect-chemical and electrochemical processes in metal-insulator-metal diodes.

In this work, we will discuss the challenges to studying these systems in general, and TiO2 in particular, as well as how this leads to vast discrepancies in the literature and competing claims over existential properties of the resistive switching filaments, from the size of switching regions, to their composition, and to their mechanism of action. To address these issues, we will introduce and discuss Electron Beam Induced Current Microscopy as a means of filament observation which provides strong, stateful contrast coupled to the resistivity of the switching region. Through comparison to Monte Carlo Simulations, we will correlate energy dependent scattering of electrons in different layers of a device stack to the generated current response so as to develop a comprehensive theory of contrast and image formation.

With a solid framework of stateful characterization in place, we will characterize the dramatic device changes in response to manufacturing variations, including subtle shifts in the composition as well as the addition of sputtered Al2O3 barrier layers. By direct observation, we will observe expansion of the filament diameter due to division of the resistive switching domain. A linear dependence between current and filament size will be measured and shown to be tunable with the barrier thickness. Dramatic reductions in operation current observed due to the Al2O3 will be definitively attributed to the suppression of surrounding damaged regions around the device.

With new insight towards the underlying physics of the switching and failure mechanisms of the devices, we will develop individual resistive switches with sufficiently improved performance due to parasitic leakage suppression to be integrated in dense, passive crossbar arrays. These crossbars demonstrated the first ever passive and integrated operations relevant to Neuromorphic computing, logic-in-memory, and system security.