The aim of this project is to simulate the hardware of an integrated photonic circuit to be used as a machine learning accelerator. The design consists of a programmable mesh of Mach-Zehnder Interferometers that can be tuned under an applied voltage to perform matrix-vector multiplication on optically encoded data. Such optical computation is extremely fast, being limited only by the speed of light in the integrated waveguides and the bandwidth of the internal phase-shifters. Additionally, the energy consumption of the hardware can be comparable to or lower than that of traditional GPUs per multiply-accumulate operation. These two benefits position optical machine learning accelerators as a promising alternative to traditional machine learning computation, with the drawback of challenges in scalability and manufacturing. In this project, we have trained a convolutional neural network on the MNIST handwritten-digit dataset and are using the Lumerical Interconnect software to simulate the optical computation involved in identifying the digit. 

Team Members:

Christian Daniel

Nathaniel Irwin

Om Joshi

Alex Meeler

Patrick Wang