Memristive oxide thin films for neural networks and artificial neurons
Memristive oxide thin films are becoming a focus of interest for application in neuromorphic computing, specifically as artificial neurons and in artificial neural networks (ANNs). Two-terminal metal-oxide-metal (MOM) devices based on niobium oxide (NbOx) are of specific interest due to their reliable threshold switching characteristics and ability to generate relaxation oscillations, or spiking. Spiking behaviour of such a device shows promise for emulation of biological neuron spiking. Additionally, nanoporous oxide films have also displayed resistive switching properties and are beginning to see application as memory devices and ANNs. The nanoscale structure and nonlinearity of nanoporous oxide films show promise for fabrication of a high-density ANN, specifically as the reservoir of a reservoir computing (RC) network.
This thesis explores the spiking behaviour of a NbOx threshold switching device and the properties of the spiking close to the initial threshold of spike generation. Close to the threshold, the device circuit exhibits stochastic spiking wholly driven by electrical noise in the circuit. With increasing input voltage, the spiking transitions into a deterministic, self-sustained form. The behavioural transition is marked by a transition of the scaling of the average interspike interval (ISI) with input voltage, which changes from a logarithmic form to a power law with exponent -½. Comparison of experimental behaviour to a simulated circuit sheds light on the physical mechanisms underpinning the switching behaviour. Parallels are drawn between NbOx device spiking and the behaviour of a type-I neuron model.
The next study focuses on fabrication of nanoporous memristive oxide thin films of Nb2O5 for application in RC. Nanoporous films are fabricated via electrochemical anodization and their electrical behaviour is characterised. Nb2O5 films form a disordered structure of nanopores 8 to 70nm in diameter. The Nb2O5 films display nonlinearity and volatile resistive switching behaviour under voltage input. The nonlinear response and high-density of the nanoporous structure confirm the candidacy of such films for application as RC reservoirs.
Nanoporous Nb2O5 devices are studied further and applied as the reservoir in a RC network. The films display co-existence of volatile and non-volatile resistive switching controlled by applied voltage. This is attributed to a gradient of oxygen vacancies across the film. The film composition is confirmed to be Nb2O5 using x-ray photoelectron spectroscopy (XPS). Nb2O5 devices display the nonlinearity and short-term memory behaviour necessary for application as a reservoir. Using a single nanoporous device and four virtual nodes as a reservoir, image recognition tasks are performed using binary images of digits 0 to 9. The network is trained using a logistic regression method and 100% recognition accuracy is obtained using only 11 training iterations and a single training data set. As such, a RC network using nanoporous Nb2O5 is shown to achieve high recognition accuracy during classification tasks whilst maintaining a low computational cost.
Rights holder© Ben Andrew Johnson
NotesA Doctoral Thesis. Submitted in partial fulfilment of the requirements for the award of the degree of Doctor of Philosophy of Loughborough University.
Supervisor(s)Pavel Borisov ; Michael Cropper
This submission includes a signed certificate in addition to the thesis file(s)
- I have submitted a signed certificate