Analysis of silicon photonics neuromorphic networks

Artificial neural networks find applications in electronics, informatics, medical purposes and many more nowadays. As the computational demand increases, a well know drawback of classic electronics is its high power consumption. In the last few years many solutions have been proposed to overcome this problem, one of the most promising being a hybrid design which uses a silicon photonic circuit as weight matrix block. The use of photonics is encouraging thanks to its low power consumption and high speed. In this work I focused on a 3x3 silicon photonic circuit made of Mach-Zehnder interferometers. It was originally designed as optical fiber switch and later adapted to neuromorphic computing by a Dutch research group. The goal of my thesis is to gain a better understanding of the thermal effects that occurs because of the Joule heating of the MZIs heaters, and study how this can affect the circuit behavior due to thermal crosstalk. I developed a MATLAB model able to predict the optical power at the device output as a function of the voltages applied to the switches and the optical input signals. Many optical components are involved in a circuit like the proposed one, such as MultiMode Interferometers for coupling, Mach-Zehnder Interferometers as weights and cross waveguides for interconnections: they have all been simulated using numerical techniques. FEM simulations were used to properly introduce thermal cross-talk. A very good match between my theoretical model and experimental measurements was obtained. Finally, the model is used to predict the circuit behavior and train a neural network later used for inverse prediction of the voltages required to obtain a requested output power.