基于孤子晶体光频梳交织的密集波分复用技术

Objective On-chip Kerr microcombs have become a key enabler for large-scale photonic signal processing, offering programmable filtering, channelisation, convolutional acceleration, integral/differential computing, and wide-band beam-forming through the time-wavelength interleaving effect. However, the usable channel count of a single microring is intrinsically limited to fewer than 100 by cavity dispersion, and the spectral energy is concentrated around the pump, leaving most of the C-band underutilised. Consequently, current comb-driven systems face a fundamental trade-off between spectral span and parallel channel density, which directly restricts the number of optical taps, convolutional kernels, or beam-forming delays that can be implemented simultaneously. To overcome this bottleneck, we propose and demonstrate a dual-interleaved soliton-crystal microcomb architecture that integrates two isomorphic, homogeneous silicon-nitride microrings on the same chip. By thermally detuning their centre wavelengths while maintaining identical free-spectral ranges and dense spectral overlap, the scheme doubles the usable line count to 200 across a >80 nm window (1520?1600 nm) while preserving a flat, high-signal-to-noise (>40 dB) power envelope. This work aims to provide a scalable, foundry-compatible pathway toward ultra-dense wavelength-division multiplexing and massively parallel photonic computing without increasing footprint, pump power, or feedback complexity.

Methods We began by optimizing the microring geometry to achieve anomalous dispersion in the C-band, targeting a 100 GHz free-spectral range and a loaded quality factor near 1×106. Linear dispersion, resonance wavelength, and intrinsic Q were extracted with sub-picometre accuracy using a swept-frequency interferometer referenced to an H¹³C¹⁴N gas cell and a 40-m unequal-arm Mach-Zehnder interferometer (MZI). Using these measured parameters, we solved the normalized Lugiato-Lefever equation via a symmetric split-step Fourier method, treating mode-crossing as a local perturbation to the second-order dispersion. Simulations mapped the (power, detuning) parameter space to identify the soliton-crystal regime and to predict the precise pump-to-crossing distance that yields optimal line-to-line flatness. Two microring chips fabricated in the same process were selected. Each chip was mounted in an alloy mold, and its temperature was controlled using a thermo-electric cooler. A continuous-wave tunable laser, amplified to 2 W by an erbium-doped fibre amplifier, served as the pump source. Through thermal tuning, the soliton-crystal combs from ring 1 and ring 2 were precisely aligned. Following power balancing, the two outputs were interleaved using a 3 dB fibre coupler, producing a 200-line, 80-nm-wide, and power-flat spectrum for subsequent convolution demonstrations.

Results and Discussions Interleaving the two thermally tuned soliton-crystal microcombs yielded a 200-line spectrum spanning 1520?1600 nm, thereby doubling the channel density compared with a single comb. To verify the utility of these additional channels, we configured a photonic convolution processor operating on 90 C-band lines simultaneously. Ten 3×3 kernels were encoded by a spectral shaper, with each line modulated at 110 Gbaud, achieving an aggregate compute rate of 19.8 TOPS.

Conclusions This work proposes and experimentally validates an interleaved wavelength-division multiplexing architecture based on dual-microring Kerr combs. The system delivers 200 parallel wavelength channels with wide spectral coverage, doubling the channel count of conventional single-comb sources. By leveraging this platform, we implemented a photonic convolutional accelerator that concurrently processes ten 3×3 kernels across 90 C-band wavelength channels, achieving a measured processing rate of 19.8 TOPS. Scaling to the full C+L band increases the throughput to 44 TOPS without increasing footprint or pump power, offering a readily extendable, chip-level solution for high-throughput, low-crosstalk optical intelligent signal processing.