The training and usage of artificial neural networks forms an increasingly large fraction of total compute. Traditional Von-Neumann digital hardware, however, struggles to efficiently emulate neural computation models. Furthermore, it no longer offers guaranteed improvements due to the slowdown of Moore's Law. As such, there is renewed interest in developing neuron-like (neuromorphic) hardware. Neuromorphic photonics, the emulation of neural compute using photonic components, is particularly attractive due to the inherent suitability of light for communication. Photonic integrated circuits (PICs), and specifically silicon photonics, promises the manufacturing scale for its realization.

My thesis consists of a collection of results related to neuromorphic photonics in silicon. It begins with a short survey of "standard" silicon photonic neurons, explored through a series of system-level results based on designs and packages I co-developed as part of my PhD work. These "standard" systems leverage typical photonic components borrowed from telecommunications. The reconfigurable neuronal synapses are formed from thermallly-tuned on-chip filters, which is power inefficient and requires external memory and dedicated drivers for each device. The neuron activations use conventional on-chip detectors and modulators, which in pure silicon photonic platforms without active electronics introduces unacceptable gain-bandwidth tradeoffs under the requirement of neural cascadability. This basic architecture and these challenges set the stage for my main original contributions, which consist of experiments and simulations exploring novel photochemical and optoelectronic silicon photonic devices with built-in neuromorphic functionality.

The first device presented is a microring resonator weight configured by the photochemical tuning of a photochromic cladding, an example of on-chip analog memory. To our knowledge, this constitutes one of the first demonstration of commercial-grade, backend compatible light molecule deposition enhancing a silicon PIC. We show interesting properties of this all-optical memory for a silicon photonics platform, including nonvolatility, low-loss in the optical C-band, and first-order photokinetics of the photoconversion leading to continuous, bidirectional, scalable actuation. The limitations of this device are also discussed, namely stability and speed. As a minor aside, we show through computation that, operating in the visible band, this device fulfills the "fingerprints" for a generic memristive device, which could have implications for online learning. The second device is a capacitively-driven silicon microring resonator colocated with a capacitive analog memory. Unlike the previous example, this memory is volatile, but offers higher reconfiguration speeds. The third example is a hybrid detector-modulator integrating and thresholding device which sidesteps the high-speed transimpedance issue of standard silicon photonic neurons in conventional silicon photonic platforms. Throughout, contributions to open-source PIC design, layout, and verification software will be highlighted.

Adviser: Paul Prucnal