Advanced Designs of Silicon Mirroring Resonators (MRR) in Wavelength Division Multiplexing (WDM) Weight Banks

Abstract

Over the past few decades, artificial neural network (ANN) architectures has been continually investigated for both academic prototype and commercial use— partially due to increasing demand for adaptive high-volume real-time processing and breakthroughs in microelectronic manufacturing fields. The research detailed in this thesis pertains to the development and research of interconnection in a spiking neural network, which is an optical variety of neuromorphic architecture. Massively parallel and enormously fast, optical technology offers promising analog interconnects whose optical signal transmissions could be multiplexed in time, space and wavelength domains. Microring resonators (MRRs) are common photonic circuit elements ubiquitous in WDM systems. In silicon photonic circuits, they are easily manufactured by bending a waveguide into a loop and o↵er the wavelength selectivity to perform tunable WDM functions.The research reported in this thesis involves advanced designs of silicon MRRs in WDM weight banks with particular interest on increasing robustness against fabrication variations. In the first stage of project, we achieved calibrated, complementary and continuous tuning of MRRs to build configurable weight banks. Next, we investigated and quantified coherent optical cross-talk in weight banks by designing new experimental device structures and developing a parameterized modeling system. Combined experimental and simulation efforts resulted in a new MRR design metric that demonstrates a tradeoff between channel space density and power penalty in weight banks. Using the same methodology, we demonstrated that 2-pole MRR devices are superior to 1-pole MRR devices in regards to this tradeoff. Lastly, our efforts concluded that a 2-pole MRR device with an inter-ring gap of 0.3nm to be the most robust against fabrication variation.