Development of Optically Communicating Nanowire-based III-V Devices: Optical broadcasting for artificial neural networks

This thesis presents the development and study of a flexible and Silicon-chip compatible integration of III–V nanowire (NW) optoelectronic components in network circuits, where communication is done via the broadcasting of light signals. This includes an investigation into how neural processing behaviours can be emulated using on-chip photonic circuits. Both experimental works that demonstrate the first on-chip communication between nano-optoelectronic components, and modelling of the optical field emission from such components are presented.
Through the development and characterisation of indium phosphide (InP) NW device pairs, optical communication across separations of up to 3 micron in air is experimentally demonstrated, which is the first step towards larger broadcasting networks. We estimate that significantly larger communication distances are possible through component performance optimisation, and by embedded waveguide materials. With yields up to 80%, the presented fabrication pipeline allows scalable assembly of heterogeneous NWs, precisely positioned in designed assemblies of tens to hundreds. We also explore the integration of photochromic dye as an all-optical localised memory medium in between the optoelectronic III-V NWs. The optically responsive behaviour of the prototype devices suggests their potential as synaptic memory elements in nanowire-based neurons.
Simulated work complements these developments by exploring how the NW optoelectronic components can form the basis of various broadcasting networks. In randomly oriented assemblies placed on a hexagonal grid, simulated InP emitter-receiver nodes were used a trainable echo-state network (ESNs), capable of mimicking a chaotic time series after minimal training. Both temporal and amplitude components of the time-series were preserved in larger networks, with smaller networks mostly replicating the frequency. The further use of directional Yagi-Uda metal nanoantenna, with added asymmetric director components, enables the directing and splitting of light from NW emitters, allowing for tuneable weights between them. This can be used to create compact layers of neural nodes with complex patterns of inhibiting and exciting weights between the nodes.
Together, these results demonstrate the feasibility of integrating nanophotonic components into either geometrically structured or trainable neural networks, building towards circuits that can for example mimic important navigation functions of the insect brain.