Silicon Loop-Type Multimode Waveguide Structure With Fan-Out Output for Photonic Reservoir Computing

• Published in: Journal of lightwave technology 2024-10, Vol.42 (20), p.7321-7329
• Main Authors: Heinsalu, Siim, Kan, Takashi, Oshima, Hirotaka, Tanaka, Hideaki, Suzuki, Masatoshi, Utaka, Katsuyuki
• Format: Article
• Language: English
• Subjects: Computation; Configuration management; Coupling coefficients; Couplings; fan-out configuration; Fanout; Impulse response; loop waveguide; multimode waveguides; NARMAm task; Nodes; Photonics; Reservoir computing; Reservoirs; Silicon; silicon photonics; Special issues and sections; Structural design; Task analysis; Training; Waveguides
• ISSN: 0733-8724; 1558-2213
• DOI: 10.1109/JLT.2024.3421522

Silicon loop-type multimode waveguide structure with fan-out output was proposed as an efficient configuration for photonic reservoir computing (RC). The device aimed to enhance node interaction between spatial and temporal-nodes through its loop-like configuration and the use of multimode waveguides. The structural design, including novel triangular configurations as input and output coupling regions, prioritized low loss and adjustable coupling coefficient, respectively. Fundamental characteristics necessary for RC were evaluated such as spatial effects through field profile imaging and temporal effects via output pulse observations. Experimental validation included speckle observation with near-field imaging, mode mixing with transmission spectra, and impulse response measurements. In our specific case, the final device employed a multimode waveguide with a width of 25 μm and a loop length of 15 mm, featuring 65 spatial nodes and 13 temporal nodes. Fan-out output was utilized to measure higher-order modes without significant losses. Our experimental findings showcased that a single pulse exhibited a spreading factor of 2.35 and could circulate up to four times. Additionally, the longest anticipated memory duration was approximately 800 ps. Also, the RC performances were evaluated in terms of memory capacity (MC) of about 11 and NARMA3 task with its NMSE of about 4 × 10 −3 , which seem superior to the previous result. Furthermore, the used parameters were also confirmed to be scalable, suggesting the potential for achieving expanding node counts in future work for higher RC performances.