Fabrication and characterisation of artificial neurons based on memristors

Volatile memristors hold significant potential in the field of neuromorphic computing, specifically in terms of applications as artificial neurons in neural networks which replicate the functions of biological systems. Operating through the phenomenon of resistive switching, volatile memristive devices transition from high to low resistance states above a characteristic threshold voltage, before returning to this high resistance state once applied voltage is removed. Two types of volatile devices have been studied in this work, each consisting of contrasting materials and switching mechanisms. The first of which is the diffusive memristor, possessing switching layers containing Ag nanoparticles suspended in the matrix of silicon dioxide (SiO2), sandwiched between two metallic electrodes. Also characterised are memristors with niobium oxide (NbOx) switching layers, also located between two metallic electrodes.<br>The technique of electrochemical impedance spectroscopy (EIS) has been used to characterise diffusive memristors, in the first instance of such experiments being applied to volatile resistive switching devices. For varying dc bias voltages, the real and imaginary components of impedance have been measured across a range of<br>ac frequencies. Resultant impedance spectra have been fitted to equivalent circuits, where the parameters of such circuits are quantitatively analysed for varying dc bias in order to assess the voltage dependence of both device resistance and capacitance. Fitted values of resistance are in strong agreement with differential resistance values<br>taken from I-V curves. Meanwhile, the fitted capacitance in one diffusive device was observed to increase significantly at higher voltages as it approached a conductive state, attributed to the growing Ag filament inside the device.<br>Both types of characterised memristor possess certain neuromorphic properties, such as the ability to demonstrate spikes in voltage in response to a dc input signal. Of particular interest to the field of meromorphic computing, the capabilities of these devices to demonstrate inhibitory and excitatory spiking have been investigated, defined as decreases and increases in spike frequency over time, respectively. These processes are fundamental in the operation of human brains, which can be described as a large system of neurons containing both inhibitory and excitatory parts. Experimentally, by connecting an additional parallel RC connection in series to a device and adjusting relevant circuit parameters, a memristor’s spiking behaviour can be altered to mimic either of these processes. The charging process of this additional capacitor causes voltage to drop across a memristor, creating a spiking response which changes in frequency over time. Initially-high frequency spiking can stop entirely once the capacitor causes voltage to drop below a required threshold, in an example of inhibitory behaviour. Inversely, a higher voltage input can be too large to initially cause voltage spikes in a device; though the charging capacitor can once again cause this voltage to drop across the memristor, allowing spike frequency to increase over time, in an example of excitatory behaviour.<br>Also discussed in this thesis are the capabilities of diffusive devices of operating as part of multiple-memristor systems. Firstly, two devices have been coupled through both resistive and capacitive coupling in order to investigate the synchronisation of spiking between the two memristors. Furthermore, the effect on synchronisation<br>strength of varying the relevant coupling parameter of either resistance or capacitance has been measured for such a system. In resistive coupling, it was determined that a lower coupling parameter provided the greatest level of synchronisation between the two devices, attributed to a path of lower impedance allowing stronger spiking signals to reach each memristor. Oppositely, for capacitive coupling, it was the largest coupling parameter measured which created the strongest synchronisation; though this can too be attributed to a path of lower impedance existing at this parameter.<br>Additionally, in a separate experiment, two devices each connected to their own individual input voltage were both joined to a third memristor which was only supplied by the outputs of the two previous devices. At varying input voltage combinations, the presence of spiking was recorded for each of the three devices, revealing trends in spiking behaviour for comparable voltages. Such measurements are compared to simulations, showing a relatively good agreement in spiking responses. At varying combinations of input voltages, different regimes of spiking were created in which one, two, or all three devices demonstrated spiking behaviour for a given pair of input signals. At higher voltages, where all three devices spike simultaneously, the system becomes a physical realisation of an integrate-and-fire neuron. The results presented are directly applicable to the field of neuromorphic computing, yielding information on how memristive devices can be used as artificial neurons to mimic the behaviours of biological networks.<br>