UC San Diego

Reservoir computing (RC) is a computational framework derived from Recurrent Neural Network (RNN), demonstrated advantage in time-series prediction and signal processing. Unlike traditional machine learning methods that rely on iterative weight optimization across the entire network, RC simplifies the training process by focusing only on a linear readout layer. This unique feature makes RC particularly attractive for applications requiring rapid adaptability, low computational overhead, and real-time data processing.The heart of RC is the ‘reservoir’, which represents a high complexity, nonlinear system with short-term memory. The reservoir itself is typically a fixed structure, which can be treated as a ‘black box’ since the training happens outside the reservoir. So it is possible for a physical system to be the reservoir, lead to the research in physical reservoir computing (PRC), which offer possibilities for ultra-high efficiency computation. A wide range of physical systems can serve as reservoir, among which, photonic systems stand out for their exceptional data transmitting speed, high bandwidth, and energy efficiency. In this dissertation, we will present two PRC researches in photonic platform based on two different directions: 1. New physical reservoir system development. We study a novel optofluidic silicon photonics system with self-induced phase change effect which relies on the coupling between geometric changes of thin liquid film and optical properties of photonic mode in waveguide, both theoretically and experimentally. The thin liquid film can operate as a nonlinear actuator and memory element, both residing at the same compact spatial region. The resulting dynamics allows to implement RC at spatial region which is approximately five orders of magnitude smaller compared to state-of-the-art experimental liquid-based system. 2. Performance enhancement in existing PRC system. We investigate the potential of using Polarization Division Multiplexing (PDM) to enhance the performance of the photonic-based PRC system. Based on a Hybrid-Photonic-Electronic RC platform, we validate through simulation and experiments that PDM improves system performance by increasing the dimensions of the photonic reservoir. The results indicate decent performance improvements with minimum increase in system footprint, highlighting the potential of PDM as a scalable and efficient tool for photonic-based PRC systems.