Metal oxide memristors for neuromorphic electronics

• Main Author: Ng, Timothy Si En
• Other Authors: Nripan Mathews
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Materials
• Online Access: https://hdl.handle.net/10356/155068
• Institution: Nanyang Technological University

Neuromorphic electronics aim to emulate the functionalities of the brain and enable the next generation of power efficient devices for artificial intelligence applications. Neuromorphic devices emulate various neural features including synaptic plasticity- which is the brain’s ability to re-wire itself during training and is credited for the brain’s efficiency in processing information and learning. Although silicon CMOS based approaches exist, efficient implementation of neuromorphic behavior necessitates the development of memristive devices based on novel materials and device architectures. Present memristive approaches for neuromorphic electronics utilize conventional two- terminal oxide memristors which operate based on filamentary formation and rupture. These conventional devices, although suitable for non-volatile memory applications are less suited for neuromorphic electronics since they do not natively demonstrate multiple states and critically show poor control over the temporal response. Essentially, these devices fall behind in terms of the required analogue programmability due to its stochastic nature. As memristors are devices that encode information in conductivity levels, it is important to study the properties not only of the bulk materials but also the interfaces to induce richer neuro-emulative behavior. Alternative programming modes and device structures which go beyond electrical biasing of 2 electrodes need to be explored. The overarching aim of this thesis is to explore such approaches to improve memristive device performances by tuning the number of states and temporal response. This can only be achieved through a mechanistic understanding of the charge transfer, charge transport and chemical changes in the oxide memristor device. The work presented in the thesis was able to address various challenges in the area of memristive devices for neuromorphic electronics by – (1) inducing multiple states in a memristor device by control over filament formation, (2) tuning the temporal (transient) response of the devices by utilization of electrical and optical stimulation and (3) incorporating simultaneous state and temporal control via a third electrode and optical input. Based on these approaches, devices with multi-state programmability and temporal tunability has been shown. Lastly, these properties were utilized to emulate functionalities inspired by biological systems demonstrating synaptic Abstract plasticity, the phenomenon of inhibition in memory formation and sensory adaptation to intensities similar to the eye. In the first work that focuses on filamentary memristors - the controllability of intermediate switching is targeted. Resistive switching is often explained by the formation and rupturing of conductive filaments comprised of oxygen vacancies. With an ion-blocking electrode impeding the oxidative process, the suppressed anodic reaction can lead to poor controllability of the resistive switching process. By employing electrodes of high reduction potential or oxygen affinity, the memristor is shown to have improved multi-state programmability. Through spectroscopic analysis, the formation of Sn(0) species during the resistive switching process was verified ex situ. With multi-state programmability, synaptic functionalities such as short-term plasticity, long-term plasticity and spike-timing dependent plasticity are emulated. While intermediate states can be achieved, the level of temporal response in filamentary switching is insufficient for neuromorphic application. Temporal response similar to a biological synapse was targeted using light as an additional means to modulate the conductance of the memristive systems. In the second work that focuses on photo-modulation – the inhibition of long-term weight changes is explored. Here, a photomemristor device based on the Schottky ITO/SnOx interface is fabricated. With a low-temperature low-oxidative thermal atomic layer deposition process, the SnOx thin film with high oxygen vacancies acts as the active switching layer. Through impedance measurements, the photomodulation mechanism is compared with the interfacial switching process. The switching behaviour of optical devices can be explained by the de-trapping of electrons at the oxide-oxide interface via photo-excitation. However, the same interface is also responsible for electrical based memristor switching. By applying electrical pulses before the optical pulses, we have shown that the memory retention of optical trainings can be tuned on demand via inhibition and facilitation. Lastly, with electrical pre-exposure, a self-filtering function known as latent inhibition can be demonstrated in associative learning by the synapse. Abstract Finally, in the third work a system which allows for simultaneous control of states and temporal decay was demonstrated. While a strong dual mode coupling can be demonstrated in photomemristors, they lack selectivity in terms of photoresponse. With a third electrode as a gating option, doping/de-doping of the electrochromic MoO3 semiconductor with Li+ can be used to tune the optoelectronic properties and hence the photoresponse. Through spectroscopic analysis, the alteration in band structure of MoO3 can be verified. This explains how both the optical and charge transport properties can be modulated with electrochemical doping/de-doping. While electrochemical transistors have been explored as synaptic devices, the tunability of optoelectronic properties is not yet utilized. Here, with reversible electrochromic switching, multiple state programmability and temporal response tunability can be demonstrated in the electrochromic transistor device. Lastly, with the tunability in photoresponsivity, adaptation features such as scotopic and photopic vision in the eye can be emulated. The thesis concludes with how further improvement on analog properties and the emulation of neuronal features can be explored. Lastly, with the development of functional memristors in processing and sensing, the opportunities in memristor-enabled computing and sensing platforms are discussed.