As artificial intelligence (AI) and data-intensive applications accelerate, they are outpacing the capabilities of complementary metal-oxide-semiconductor (CMOS) technology, sustained for decades by Moore's Law. However, as Moore's Law approaches physical and practical barriers, there is a shift in research interest from conventional Von-Neumann architectures toward brain-inspired neuromorphic computing (NC). This paradigm utilises alternative hardware technologies that emulate the biological brain's remarkable efficiency and adaptive capabilities across temporal scales, from short-term memory (STM) to long-term memory (LTM). The brain's neural networks, overseen by synaptic plasticity and inherent neuronal dynamics allow for excellent information encoding and adaptation to various environmental inputs, from instantaneous sensory responses to long-term learning.This thesis investigates TiOx-based memristors as a hardware element for NC, utilising their synaptic-like characteristics to mimic STM and LTM behaviours. Memristors exhibit resistance-dependent memory, which offers a compact alternative to CMOS with tuneable synaptic potentiation and depression across timescales from nanoseconds to years. Despite promising demonstrations in memristor-based reservoir computing (RC) systems, the consideration of temporal dynamics is often underexplored. This work investigates both volatile and non-volatile TiOx-based memristors under pulse modulation, focusing on exploiting the memristors' intrinsic temporal decay and multi-parameter states to demonstrate a concept for improving the RCâs robustness against input noise variations. The fabricated TiOx-based memristors exhibit volatile short-term decay on timescales ranging from microseconds to milliseconds, while memristors with non-volatile long-term retention operate in the millisecond timescale.The sputtering of a Ti target in a reactive oxygen-argon mixture produces a mixed-phase material consisting of TiO2 and Ti2O3. By altering the oxygen flow rate, the relative ratio of TiO2 and Ti2O3 in the deposited thin film for memristors can be modified. These TiO2-rich and TiO2-deficient phases were investigated using X-ray photoelectron spectroscopy (XPS) analysis and pulse modulation. It was found that a reduced oxygen flow rate during sputtering introduces higher oxygen deficiency, resulting in a sub-stoichiometric TiO2-deficient composition that promotes volatile short-term memory. In contrast, a higher oxygen flow rate minimises deficiency by allowing more complete oxidation, producing a near-stoichiometric TiO2-rich film that supports non-volatile long-term memory.The volatile TiO2-deficient memristors achieved an accuracy of approximately 90% under ideal conditions for the MNIST image classification task, and 66% accuracy even after an approximately 1.7
Alexander-Hanyu Wang (Thu,) studied this question.