Modeling Photonic Integrated Circuits for Optical Computing using Machine Learning

The past decade has witnessed an artificial intelligence revolution leading to the widespread integration of artificial intelligence in various industries and daily life. Yet, traditional electronic computers are projected to fall short of meeting increasing computational needs to power artificial intelligence applications. To this end, photonic integrated circuits stand as a promising candidate to enable innovation in computing in the form of optical computers, which hold the potential to become both faster and more energy-efficient than their electronic counterparts. However, modeling and controlling such circuits has proven to be a challenging task and remains to be an active research area. Machine learning methods hold great promise for addressing various challenges associated with modeling photonic chips for optical computing. This thesis contributes to the state of the art in modeling photonic integrated circuits for optical computing by investigating several machine learning approaches. Firstly, offline training of machine learning models describing Mach-Zehnder interferometer-based optical matrix multipliers is studied. The performance of both analytical and data-driven models is evaluated using experimental data collected from an optical matrix multiplier chip suffering from thermal crosstalk. The results indicate that neural network models significantly enhance modeling accuracy and are even able to model the spectral response of the chip. To address experimental data scarcity issues for training such models, a transfer learning approach is proposed. The investigation showcases that significant reductions in modeling errors can be achieved by this method when training data is limited. Furthermore, a well-regularized ensemble of neural network models is shown to further improve modeling performance. Finally, the use of machine learning methods is extended to the case of thermal crosstalk compensation for programmable photonic processors. A microring resonator implemented by a photonic processor is used to investigate different approaches for modeling the spectral shifts due to thermal crosstalk. Linear regression models are found to be the best at modeling and compensating for thermal crosstalk, while experimentally fitted analytical models also achieve impressive results. In addition, the analytical models trained using one microring resonator can also be used to model the effect of thermal crosstalk for another resonator located at a different location on the same chip.