Study of the Nonlinearities of Microrings in Silicon Photonics for Neuromorphic Computing

The objective of this master's thesis is to develop innovative tools that exploit the nonlinear properties of optical micro-ring resonators, integrated in silicon photonics platform, to create a neural network capable of solving complex tasks such as image and voice recognition, time series prediction, using the time delay reservoir computing (RC) technique. These micro-ring resonators are part of a photonic integrated circuit in silicon that can be fabricated in the same CMOS foundry of electronic integrated circuits. Photonic implementations of reservoir computing are promising applications due to their ability to achieve very high bandwidth and to exploit the intrinsic parallelism of optical signals. My thesis has three main objectives: to develop a compact fully optical RC model based on a silicon optical micro-ring resonator to classify datasets into high-speed clusters; to analyze and understand the nonlinear optical effects in micro-rings during these processes; to create an experimental setup to validate this RC method. The experimental work is carried out at the Interdepartmental Center Photonext at PoliTO. To achieve these goals, I used a silicon micro-ring resonator in add-drop configuration in a photonic integrated circuit, designed at PoliTo and fabricated in IMEC (Belgium). To demonstrate how the nonlinearities of micro-ring resonators can be exploited within a neural network, I reproduced the experiments and the theory presented in the article “Reservoir computing based on a silicon microring and time multiplexing for binary and analog operations” in Pavesi, L. et al. Scientific Reports 2021, 11, 15642. I began my study with a compact theoretical model, developed by Dr. Marco Novarese in his PhD thesis that describes the impact of nonlinear (NL) effects on the performance of ring resonators. In neural networks, the input is mapped into a higher-dimensional space and processed by nonlinear functions while it passes from one layer to the next. I replicate this functionality using the nonlinear response of the micro-ring resonator. In this context, I addressed classification problems such as the delayed XOR task and the recognition of the Iris species. To validate my approach, I developed a pump-probe experimental setup. I encoded the input into the temporal evolution of the intensity of a high-power pump laser and I used another low-power laser as a probe. I centered these two lasers at two adjacent resonant frequencies. I modulated the laser input using a Mach Zehnder through a wave, generated by an AWG, with the same trend as the coded input. The pump is then amplified and combined with the probe through a coupler 50%50% to finally enter in the micro-ring. The pump induces nonlinear phenomena in the device, and therefore, I used the probe to detect the induced variations in the microring. At the output of the microring, I filtered the probe's wavelength and directed it through an avalanche photodetector to an oscilloscope for data acquisition and analysis. Using MATLAB, I sampled the virtual nodes and trained the network, achieving classification rates above 90%. By testing the XOR dataset, I demonstrated the presence of memory and showed that sampling the probe instead of the pump improves the classification accuracy. Additionally, I found that as the bitrate increases, the accuracy decreases. This research opens new perspectives for the implementation of high-speed, low-energy optical neural networks, leveraging the potential of silicon photonics.