Photonic Time-delay Reservoir Computing based on Silicon Microring Resonators

The progress of artificial intelligence over the past decade has transformed fields like healthcare, finance, transportation, and communications, increasingly impacting daily life, yet the growing computational demands of large neural networks and data centers expose the limitations of traditional electronic computing. Physical constraints, such as those imposed by quantum mechanics, are bringing Moore’s Law to its limits. Beyond computational power, other issues, such as high energy consumption, heat dissipation, and data latency, highlight the need for alternative computing frameworks. Optical computing, harnessing the promising benefits of light such as higher speed, energy efficiency, and parallelism, has emerged as a promising solution. However, large deep neural networks are not well-suited for photonic implementation, consequently, driving interest in alternative schemes with minimal hardware requirements such as time-delay reservoir computing.

This thesis investigates photonic time-delay reservoir computing exploiting two-photon absorption nonlinear effects in silicon microring resonators. A numerical analysis identifies three distinct performance regions correlating to the transition of the microring from linear to highly nonlinear regimes in function of the power and frequency of the laser source. The optimal region achieves impressive performance across tasks by balancing memory and nonlinearity, while other regions suffer performance penalty from instability or insufficient nonlinear behavior. Adjusting time constants related to the nonlinear effects by tailoring the optical properties of the resonator is shown to enhance the flexibility of the system to exhibit optimum performance.

Additionally, the potential of exploiting wavelength division multiplexing (WDM) for multitask parallel reservoir computing in the microring resonator is explored, enabling simultaneous computation of multiple tasks through multiple optical channels. The system demonstrates strong performance at both handling multiple instances of the same task, and at addressing different applications in parallel such as classification, timeseries prediction, wireless channel equalization, and prediction of radar signals. Furthermore, the investigation showcases the combination of WDM with pre-processing of the input signal to enhance memory capacity for single-task implementations of the studied system, enabling compact, high-performance photonic circuits. Overall, this research contributes to enhance the performance, memory and parallel computing capabilities of future scalable, energy-efficient photonic computing architectures.