Neuromorphic nanophotonic systems for artificial intelligence

Student thesis: Doctoral Thesis


Over the last decade, we have witnessed an astonishing pace of development in the field of artificial intelligence (AI), followed by proliferation of AI algorithms into virtually every domain of our society. While modern AI models boast impressive performance, they also require massive amounts of energy and resources for operation. This is further fuelling the research into AI-specific, optimised computing hardware. At the same time, the remarkable energy efficiency of the brain brings an interesting question: Can we further borrow from the working principles of biological intelligence to realise a more efficient artificial intelligence? This can be considered as the main research question in the field of neuromorphic engineering. Thanks to the developments in AI and recent advancements in the field of photonics and photonic integration, research into light-powered implementations of neuromorphic hardware has recently experienced a significant uptick of interest. In such hardware, the aim is to seize some of the highly desirable properties of photonics not just for communication, but also to perform computation. Neurons in the brain frequently process information (compute) and communicate using action potentials, which are brief voltage spikes that encode information in the temporal domain. Similar dynamical behaviour can be elicited in some photonic devices, at speeds multiple orders of magnitude higher. Such devices with the capability of neuron-like spiking are of significant research interest for the field of neuromorphic photonics. Two distinct types of such excitable, spiking systems operating with optical signals are studied and investigated in this thesis. First, a vertical cavity surface emitting laser (VCSEL) can be operated under a specific set of conditions to realise a high-speed, all-optical excitable photonic neuron that operates at standard telecom wavelengths. The photonic VCSEL-neuron was dynamically characterised and various information encoding mechanisms were studied in this device. In particular, a spiking rate-coding regime of operation was experimentally demonstrated, and its viability for performing spiking domain conversion of digital images was explored. Furthermore, for the first time, a joint architecture utilising a VCSEL-neuron coupled to a photonic integrated circuit (PIC) silicon microring weight bank was experimentally demonstrated in two different functional layouts. Second, an optoelectronic (O/E/O) circuit based upon a resonant tunnelling diode (RTD) was introduced. Two different types of RTD devices were studied experimentally:a higher output power, µ-scale RTD that was RF coupled to an active photodetector and a VCSEL (this layout is referred to as a PRL node); and a simplified, photosensitive RTD with nanoscale injector that was RF coupled to a VCSEL (referred to as a nanopRL node). Hallmark excitable behaviours were studied in both devices, including excitability thresholding and refractory periods. Furthermore, a more exotic resonate and-fire dynamical behaviour was also reported in the nano-pRL device. Finally, a modular numerical model of the RTD was introduced, and various information processing methods were demonstrated using both a single RTD spiking node, as well as a perceptron-type spiking neural network with physical models of optoelectronic RTD nodes serving as artificial spiking neurons.
Date of Award9 Jan 2024
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorAntonio Hurtado (Supervisor) & Michael Strain (Supervisor)

Cite this