Reservoir Computing-assisted optical system for high-throughput applications

High-throughput photonic systems enable ultrafast data acquisition in many applications, but the large volumes of generated data create significant challenges for real-time signal processing. This thesis investigates optical approaches for high-throughput applications using machine learning-based reservoir computing (RC). As a simplified form of recurrent neural networks (RNNs), RC significantly reduces training complexity while maintaining strong capability in processing time-sequential data, making it particularly suitable for classification and prediction tasks in high-speed systems.

RC is first employed as a back-end data analysis tool for two high-throughput applications: indoor user localization and frequency hopping recognition. In the indoor localization system, RC is integrated with a photonic time-stretch framework using a 45° tilted fiber grating (TFG), enabling real-time tracking and accurate position identification of users. A proof-of-concept experiment demonstrates improved localization performance. In the frequency hopping recognition system, RC is applied in the post-processing stage to identify frequency transitions, enabling accurate detection of hopping time and duration while improving system responsiveness.

Building on these applications, an optical hardware implementation of reservoir computing is proposed and experimentally demonstrated. The system combines spectrum mixing, photonic time stretch, and a delay-based memory structure to enable efficient high-speed processing. In this design, both wavelength and time dimensions are utilized as reservoir nodes, significantly expanding the number of virtual nodes and overcoming limitations in conventional hardware reservoir implementations. Simulation studies investigate key system parameters, including node number, optical feedback strength, and the gain characteristics of the semiconductor optical amplifier (SOA). Experimental results further validate the system through waveform classification and frequency classification tasks.

To further improve processing speed and enable real-time operation, a novel optical masking scheme is introduced. By embedding the mask directly within the optical link, this approach eliminates the electronic bottleneck associated with conventional pre-masking procedures and enables fully optical preprocessing. Three optical mask implementations are investigated and experimentally evaluated. The results demonstrate improved classification accuracy and highlight the potential of all-optical masking for high-speed photonic reservoir computing systems.

Overall, this work demonstrates that optical reservoir computing provides an effective solution for high-throughput machine learning tasks, offering significant advantages in processing speed, scalability, and real-time capability. The proposed architectures and techniques contribute to the development of next-generation photonic computing systems for ultrafast signal processing and intelligent optical sensing.