Photonic neuromorphic processing with coupled spiking silicon microrings

TL;DR

The paper demonstrates that a minimal SCISSOR core, formed by three coupled silicon microring resonators, can function as a compact passive nonlinear neuron suitable for both analogue and digital reservoir computing. By exploring coupling-edge regimes and leveraging spiking and thermal bistabilities, the authors achieve state-of-the-art-like results on Iris (Acc=1) and Sonar (up to ≈0.984 analogue, ≈0.969 digital) with only 150 virtual nodes, highlighting a physics-guided pathway to high-performance on-chip neuromorphic photonics. The work elucidates how dynamical transitions at coupling edges correlate with task performance and presents practical training/tracking strategies around these operating points. Overall, SCISSOR-based photonic cores emerge as scalable, CMOS-compatible building blocks for fast, energy-efficient neuromorphic processing at the edge.

Abstract

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.

Figures (10)

• Figure 1: Schematic of an integrated silicon-photonic SCISSOR neuromorphic core for edge computing applications. Two-photon absorption induces free-carrier dispersion (FCD) and thermo-optic effects (TOE), which together govern the nonlinear transformation applied to optical input signals, reducing the complexity of subsequent electronic post-processing.
• Figure 2: Experimental characterization of a 3-ring SCISSOR under continuous wave (CW) optical injection. (a) Linear drop spectrum. (b) Experimental scheme where CW laser signal is injected at the input port of the device and the corresponding drop port is analysed to recreate a phase space relying on the Takens' theorem. (c) Estimation of the density of the phase space for multiple $\Delta\lambda$-CW Optical power injection configurations and, (d) corresponding mean value of the detected drop signal. (e-f) Examples of SCISSOR's dynamics and corresponding fast Fourier transforms (fft) under injection configurations labelled by coloured (e) and numbered (f) circles, in (c) and (d). The time traces and their fft are evaluated over 10 $\mu$s of recorded data with 2 $ns$ sample rate; for visualization, only a 2 $\mu$s portion of the temporal signal is shown in the panels. Fft panels showing the lost of harmonic peaks have a light-blue background colour.
• Figure 3: (a) Analogue and digital RC scheme (top) and their photonic-electronic implementation (bottom). Nodes belonging to the masking layer (red) are temporally encoded using a Mach Zehnder modulator, while the reservoir state (blue) unfolds in the dynamical evolution of a physical SCISSOR neuromorphic core. The output layer (black) is computed electronically. For a complete setup description, see Experimental section. (b) Optical input signal encoding the masking layer nodes as amplitude values (left) and corresponding SCISSOR nonlinear response collected from the drop port (right), from which time-multiplexed reservoir nodes are extracted in analogue (yellow) or digital (purple) form according to an arbitrary digital threshold (th).
• Figure 4: Performance on the Iris task obtained using (a) analogue and (b) digital RC for increasing values of the digital threshold (th). Top panels report the classification accuracy ($Acc$), while bottom panels show the standard deviation of the accuracy ($\sigma_{Acc}$). (c) Best-performing digital RC configurations for increasing digital thresholds (black rectangles in (b)). Left panels show the optical nonlinear transformation produced by the SCISSOR for Iris sample $\#105$, while right panels report the corresponding digital reservoir state maps. Horizontal red lines indicate the boundaries between Setosa (1-50), Versicolor (51-100) and Virginica (101-150) classes.
• Figure 5: Classification scores on the Iris task when the digital RC is trained to recognize only the Virginica specie. (a) Accuracy ($Acc$) scores obtained when setting a digital threshold (th) of 540 mV. Under an optical injection of $\Delta\lambda=-320$ pm and 5.85 mW input power (within the input waveguide) the SCISSOR obtained accuracy 1. (b) The corresponding SCISSOR optical nonlinear transformation of Flower $\#105$ (belonging to the Virginica class), and (c) map of the digital patterns obtained for the full dataset, show that only one spike allows solving the task. Horizontal red lines in the digital map indicate the boundaries between Setosa (1-50), Versicolor (51-100) and Virginica (101-150) classes.