A Generalized Silicon Photonics Device Suite: From Communications to AI

Abstract

The massive growth of telecommunications traffic has brought about the necessity for communication systems to be increasingly cheaper, faster, and more power efficient. Photonic integrated circuits (PICs) have attracted considerable interest due to their inherent advantages in meeting the ever-increasing bandwidth demand. These advantages include but are not limited to, low propagation loss, absence of Joule heating, and increased bandwidth. Certain techniques, such as wavelength division multiplexing (WDM) realized on silicon photonics platforms, pose very promising solutions for enhancing the aggregate bandwidth and enabling low-power, efficient transceivers. PICs aim to combine a vast number of optical devices in conjunction with microelectronics for optimized on-chip integration density. Aside from communications, PICs have also been rapidly gaining traction in the field of artificial intelligence (AI) accelerators. State-of-the-art AI algorithms rely on large amounts of linear algebra computations, i.e., multiply–accumulate operations (MAC). These computations can seamlessly be performed at the speed of light using meshes of photonic components such as Mach-Zehnder modulators (MZMs) and other photonic components. Even though PICs possess a wide range of advantages, one key challenge associated with them, is fabrication reproducibility, i.e., large performance variation across different locations on the silicon wafer. In this work, we experimentally demonstrate a set of silicon photonics devices aimed to be used in photonic transceivers and AI accelerators. The devices are fabricated on a state-of-the-art, 45-nm, monolithic silicon photonics platform and are specifically designed to demonstrate fabrication robustness. We develop wavelength-independent power splitters, which demonstrate broadband performance and can be tuned to attain any value of power splitting ratio (SR) depending on the application. Subsequently, we employ these splitters on a compact Mach-Zehnder Interferometer (MZI)-based cWDM filter aiming at flat-top, high crosstalk response. The devices demonstrate excellent performance stability throughout different wafer sites. Lastly, we design heaters based on micro-ring resonator-assisted MZMs (RAMZMs) as low-power, low-operating-voltage efficient optical interference units, and non-linear activation units for photonic neural networks. The fabricated passive and active devices operate in the O-band and can seamlessly be integrated into communications or AI applications. Finally, we design, train, and validate a photonic neural network based on a MZM mesh-based architecture for boolean logic applications.

Date of Award	Dec 2022
Original language	American English
Supervisor	JAIME Viegas (Supervisor)