Photonic neuromorphic processing with coupled spiking silicon microrings

Understanding the physical computing mechanisms of individual network nodes is essential for scaling neuromorphic photonic architectures. This work proposes a compact passive nonlinear photonic core based on a Side-Coupled Integrated Spaced Sequence of Resonators (SCISSOR) made of three nominally equal microrings and investigate its computing capabilities. Its nonlinearities and internal feedback enable analogue, spiking, and bistable responses that are accessed by tuning the injection power and wavelength. Implemented as a single nonlinear node in a time-multiplexed reservoir computing, the SCISSOR achieves error-free classification on the Iris dataset and accuracies above 97% on the Sonar task, using both analogue and digital reservoir representations with 150 virtual nodes. In the digital scheme, spiking dynamics naturally generate sparse reservoir states, enabling efficient classification even with a single spike. Intriguingly, optimal operating points are at the boundaries where sharp transitions in dynamical complexity and/or output power occur. In these points, the SCISSOR supports high task-performance, opening novel strategies for future on-chip training. Spiking and thermal bistabilities also participate to enhance the computational performance at low injected powers below 4 mW. These results suggest optical coupled microring resonators as effective building blocks for future edge computing and neuromorphic photonic systems.

💡 Research Summary

The authors present a compact neuromorphic photonic core based on a Side‑Coupled Integrated Spaced Sequence of Resonators (SCISSOR) consisting of three nominally identical silicon microring resonators (MRRs). By exploiting two‑photon absorption (TPA) in the rings, free‑carrier dispersion (FCD) and thermo‑optic effects (TOE) are generated on distinct timescales (nanoseconds for carriers, hundreds of nanoseconds for heat). These nonlinearities, together with the coherent feedback among the three rings, give rise to a rich set of dynamical regimes—steady‑state, self‑pulsation in a single ring, coupled self‑pulsation in two rings, and fully synchronized periodic or chaotic oscillations across all three rings. The authors map these regimes by sweeping the injection wavelength detuning (Δλ) and optical power, reconstructing a two‑dimensional phase‑space using Takens’ embedding and quantifying its density (ρ) as a proxy for dynamical complexity. High‑ρ regions correspond to strong inter‑ring coupling and complex dynamics, while low‑ρ regions indicate weak coupling or quiescent operation. Notably, abrupt changes in ρ and in the average drop‑port power occur at the “coupling‑edge” boundaries between weak and strong coupling, providing natural operating points where the system’s computational capability is maximized.

To assess computational usefulness, the SCISSOR is embedded as the nonlinear node of a time‑multiplexed reservoir computer (RC). Input data are first expanded into 150 virtual nodes via a linear masking layer; each mask value modulates the intensity of a continuous‑wave laser through a Mach‑Zehnder modulator (MZM). The modulated optical signal is injected into the SCISSOR, whose intrinsic dynamics transform the signal. The drop‑port output is detected and sampled at the same rate as the mask (48 ns per node). Two readout schemes are explored: (i) an analogue approach where the raw voltage values of all 150 virtual nodes are used directly, and (ii) a digital approach where a threshold converts each node’s response into a binary spike (1) or silence (0). The digital scheme leverages the spiking nature of the SCISSOR to produce sparse reservoir states, enabling accurate classification even when only a single spike is present in the entire reservoir state vector.

Experimental validation is performed on two benchmark classification tasks. On the Iris dataset (3 classes, 150 samples), both analogue and digital RC achieve 100 % classification accuracy, demonstrating that the SCISSOR provides sufficient nonlinear transformation power for linearly separable tasks. On the more challenging Sonar dataset (2 classes, 208 samples), the system reaches >97 % accuracy while operating at injected powers below 4 mW, highlighting the efficiency of the device. Optimal performance is observed at two distinct operating points: (a) a high‑ρ region near Δλ ≈ 20 pm and 10 mW, where the three rings are strongly coupled and exhibit chaotic‑like dynamics, and (b) a low‑ρ but spiking region near Δλ ≈ −100 pm and 4 mW, where thermal bistability and single‑ring self‑pulsation dominate. In both cases, the transition between dynamical regimes is marked by a sudden change in average output power, offering a clear experimental signature for tuning.

Key insights from the work include: (1) a minimal three‑ring photonic system can generate a spectrum of analogue, spiking, and bistable responses suitable for neuromorphic computing; (2) the boundaries between weak and strong coupling (“coupling‑edge”) act as natural sweet spots where computational performance peaks, suggesting a strategy for on‑chip training and control; (3) spiking dynamics inherently produce sparse binary reservoir states, reducing the burden on downstream electronic processing and enabling ultra‑low‑power operation; and (4) thermal and carrier‑based nonlinearities can be harnessed together to enhance performance at sub‑5 mW power levels.

The paper concludes that coupled microring resonators are promising building blocks for scalable, edge‑oriented photonic neuromorphic systems. Future directions include scaling to larger arrays of SCISSOR units, integrating on‑chip photodetectors and electronic readouts, and exploring adaptive control schemes that exploit the identified coupling‑edge operating points for online learning.