Kaoru MINOSHIMA (美濃島 薫)

Abstract

Optical reservoir computing (ORC) promises fast, energy-efficient temporal inference by harnessing the rich transient dynamics of photonic systems. Yet most ORC demonstrations still depend on fiber delay lines or camera-based spatial multiplexing, which caps the clock rate at a few tens of MSa/s and complicates monolithic integration. Here we introduce a frequency-multiplexed ORC whose nodes are the individual modes of a dissipative Kerr-soliton microcomb generated in a high-Q Si₃N₄ microresonator. The input signal is encoded as a rapid detuning modulation of the pump laser, so the intracavity dynamics of the microcomb provide both the high-dimensional nonlinear mapping and tens of nanoseconds of memory, while output weighting is realized optically with standard microring arrays. Numerical modeling with 60 comb modes provides a normalized mean-square error (NMSE) of 0.015 on the Santa Fe chaotic time-series task at 50 MSa/s and more than a tenfold reduction in symbol-error rate for nonlinear equalization (NLEQ) at 100 MSa/s. A proof-of-concept experiment using 37 measured modes also confirms the concept on the Santa Fe chaotic time-series and NLEQ benchmarks. Because both the microcomb and weighting network are fabricated by a complementary metal-oxide semiconductor (CMOS)-compatible process, the architecture offers a clear path toward compact, energy-efficient photonic processors operating at greater than 1 GSa/s, directly addressing the scalability and speed challenges of nanophotonic ORC.