Silicon microring resonator for photonic neural networks

Abstract

Photonic neural networks (PNNs) have emerged as a promising hardware platform for artificial intelligence by exploiting the high bandwidth, inherent parallelism, and low latency of optical computing. Among integrated photonic devices, silicon microring resonators, with their compact footprint and compatibility with complementary metal–oxide–semiconductor (CMOS) fabrication, have been widely investigated for implementing modulation and weighting. However, previous practical PNN systems for a specific task, such as image classification, have separated sensing and computing units, which limits the system performance. Furthermore, a practical PNN requires more than modulation and weighting. A nonlinear activation function is a necessary module for implementing multilayer neural networks, while reliable chip packaging is required to achieve a standalone integrated system. Motivated by these challenges, this thesis investigates the use of silicon microring resonators in photonic neural networks, with a focus on all-optical modulation for in-sensor imaging classification, nonlinear activation, and chip packaging implementation. First, all-optical modulation in silicon microring resonators is experimentally demonstrated and analyzed. The modulation mechanism is explained through the free carrier dispersion effect and the thermo-optic effect. A dynamic model is developed to describe the device response. Experimental measurements under different operating conditions are used to evaluate modulation performance, and an optimized operating method is proposed to enhance the modulation depth. Next, an in-sensor image classification architecture that combines all-optical modulation with PNNs is proposed, in which optical signals generated by the sensor are processed directly by a microring-based weighting network, thereby reducing unnecessary optical-electrical-optical signal conversions. Experimental dot product measurements using a microring array achieve an effective resolution of 8.1 bits. System-level evaluation based on the measured device characteristics demonstrates the feasibility of image classification using the proposed architecture. Furthermore, the investigation of microring resonators is extended to nonlinear activation functions, including both optical–electrical–optical and all-optical approaches. Finally, the chip packaging process is presented and demonstrated through a successfully packaged photonic tensor core. Overall, this thesis demonstrates the potential of silicon microring resonators as versatile building blocks for integrated PNNs. These results contribute to the development of practical, fully integrated photonic computing systems.