🔬 applied physics

Photonic neuromorphic processing with coupled spiking silicon microrings

This paper demonstrates that a compact, passive SCISSOR-based photonic node, leveraging coupled spiking and thermal bistabilities to achieve high-performance classification on benchmark datasets with low power consumption, serves as an effective building block for scalable neuromorphic photonic edge computing systems.

Original authors: Giovanni Donati, Stefano Biasi, Lorenzo Pavesi, Antonio Hurtado

Published 2026-02-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Giovanni Donati, Stefano Biasi, Lorenzo Pavesi, Antonio Hurtado

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: A Tiny Optical Brain on a Chip

Imagine you have a massive library of data (like photos of flowers or sonar recordings of rocks vs. mines) that needs to be sorted quickly. Usually, computers do this by converting light into electricity, processing it, and converting it back. This is slow and uses a lot of energy.

This paper proposes a different way: let the light do the thinking itself.

The researchers built a tiny, passive "brain" made of silicon and light. Instead of using electricity to switch transistors on and off, they use three tiny glass rings (microrings) that trap light. When light enters these rings, it bounces around, interacts with itself, and creates complex patterns. These patterns act like the "thoughts" of a neuron.

The Core Component: The SCISSOR Necklace

The main invention is called a SCISSOR (Side-Coupled Integrated Spaced Sequence of Resonators).

The Analogy: Think of a single microring as a single bell. If you ring it, it makes a simple sound.
The Innovation: The researchers linked three of these bells together in a specific chain. When you ring one, the sound vibrates through the others, creating a complex, chaotic, and beautiful harmony that a single bell could never make on its own.
How it works: They shine a laser into this chain. Depending on how much power they use and the exact color (wavelength) of the laser, the light inside the rings behaves differently. Sometimes it flows smoothly; sometimes it "spikes" (flashes on and off rapidly); sometimes it gets stuck in a loop.

The Two Ways to "Read" the Brain

The researchers tested this optical brain in two different modes, similar to how a human might read a book:

Analogue Mode (The Continuous Stream):
- Imagine reading a book where every word is a different shade of gray. The computer reads the exact brightness of the light coming out of the rings. This is very detailed and works well for simple tasks.
- Result: They used this to sort the famous "Iris flower" dataset (identifying flower types) with 100% accuracy.
Digital/Spiking Mode (The Morse Code):
- Imagine reading a book where you only care if a word is "loud" or "quiet." If the light flashes above a certain brightness, the computer counts it as a "1" (a spike). If it's dim, it's a "0".
- The Magic: Even though this method throws away a lot of detail (it's very "sparse"), it turned out to be incredibly efficient. For the flower task, they could identify the flowers using just a single flash of light.
- Result: They also tested this on a harder task: the "Sonar" dataset (distinguishing rocks from mines). They achieved over 97% accuracy using this sparse, spiking method.

The Secret Sauce: The "Edge of Chaos"

The most interesting discovery in the paper is where the system works best.

The Analogy: Think of a tightrope walker. If they stand in the middle of the rope, they are stable but boring. If they stand at the very edge, they are wobbly and chaotic.
The Finding: The researchers found that the optical brain works best when it is balanced right on the edge between stability and chaos.
- If the light is too weak, nothing happens (too stable).
- If the light is too strong or the rings are too coupled, the system goes wild and loses its ability to sort data (too chaotic).
- The Sweet Spot: When they tuned the laser to the exact point where the system is about to switch from one behavior to another (e.g., from smooth flow to flashing spikes), the computer became super smart. This is called the "coupling-edge."

Why This Matters (According to the Paper)

Efficiency: You don't need a massive computer to do this. A tiny chip with just three rings can do the work of a much larger system because the physics of the light does the heavy lifting.
Speed: Because it uses light, it avoids the slow conversion between light and electricity.
Simplicity: The system is "passive," meaning it doesn't need complex internal wiring to change its behavior. You just tune the laser (the input), and the rings naturally rearrange their "thoughts" to solve the problem.

Summary

The paper shows that by linking three tiny silicon rings together and shining a laser through them, they created a mini-computer that can sort data with incredible accuracy. The secret is tuning the laser to the exact "edge" where the light behavior changes from smooth to spiky. This allows the system to solve complex problems using very little energy and very few "spikes" of light, making it a promising building block for future, faster, and more energy-efficient computers.

Technical Summary: Photonic Neuromorphic Processing with Coupled Spiking Silicon Microrings

Problem Statement
As the computing community faces the end of Moore's Law, there is a critical need for alternative technologies to process growing data volumes efficiently, particularly for edge computing scenarios requiring low latency and high energy efficiency. Neuromorphic photonics offers a promising solution by performing nonlinear transformations and temporal processing directly in the optical domain, minimizing electro-optical conversions. While Reservoir Computing (RC) is well-suited for such tasks due to its reliance on natural physical dynamics rather than on-chip training of internal connections, current implementations face scalability challenges. Single microring resonator (MRR) nodes often lack the dynamic complexity required for challenging tasks, and scaling to large arrays of coupled MRRs in a "black box" approach leads to rapidly growing operational parameter spaces and thermal crosstalk issues. There is a pressing need to understand the computing capabilities of reduced, coupled structures to enable efficient scaling of on-chip optical neuromorphic systems.

Methodology
This work experimentally investigates a compact, passive nonlinear photonic core based on a Side-Coupled Integrated Spaced Sequence of Resonators (SCISSOR). The device consists of three nominally identical self-pulsating silicon MRRs (radius 7 µm) evanescently coupled between two bus waveguides with a symmetric gap of 180 nm. The nonlinear dynamics are driven by Two-Photon Absorption (TPA), which induces free-carrier dispersion (FCD) and thermo-optic effects (TOE), creating distinct timescales for free-carrier lifetime (1–45 ns) and thermal relaxation (60–280 ns).

The study employs a time-multiplexed RC architecture where:

Input Encoding: Data points from the Iris and Sonar datasets are linearly expanded offline into a masking layer of 150 virtual nodes. These are encoded as amplitude levels of a continuous wave (CW) tunable laser using a Mach-Zehnder Modulator (MZM).
Nonlinear Processing: The modulated optical signal is injected into the SCISSOR. The device's intrinsic dynamics (thermal bistability, self-pulsation, and spiking) perform the nonlinear transformation. A resting period (~2 µs) resets the internal memory between inputs.
Readout: The drop-port signal is photodetected and sampled. The system is evaluated in two modes:
- Analogue RC: The sampled analogue voltage values serve as reservoir states.
- Digital RC: The analogue states are thresholded to binary values (0 or 1), exploiting the device's spiking dynamics to create sparse representations.
Characterization: The SCISSOR's dynamical complexity is quantified using Takens' theorem to reconstruct phase spaces and calculate a density metric ( $\rho$ ). This is mapped against injection wavelength detuning ( $\Delta\lambda$ ) and optical power to identify operating regimes.

Key Contributions

Coupled Microring Architecture: The paper demonstrates that a minimal SCISSOR structure (three coupled rings) can access "edge of chaos" regimes and complex aperiodic dynamics arising from multi-resonator TPA-based interactions, surpassing the capabilities of single MRR nodes.
Coupling-Edge Operating Principle: A central contribution is the identification of "coupling-edge" operating points. These are boundaries between weak and strong coupling regimes where sharp transitions in dynamical complexity ( $\rho$ ) and/or output optical power occur. The authors find that these specific boundaries, rather than regions of maximum chaos, often support the highest computational performance.
Sparse Digital Representations: The work shows that the SCISSOR's spiking dynamics naturally generate sparse reservoir states in digital RC. This allows for efficient classification even with extremely low spike counts, simplifying subsequent electronic post-processing.
Bistability Utilization: The study highlights that both thermal and spiking bistabilities contribute to enhancing computational performance, particularly at low injected powers (< 4 mW).

Results

Iris Dataset: The SCISSOR-based RC system achieved error-free classification (100% accuracy) on the Iris dataset using both analogue and digital schemes with 150 virtual nodes. Notably, in the digital scheme, specific configurations allowed the task to be solved with the activation of a single temporal node (spike), demonstrating extreme efficiency.
Sonar Dataset: For the more challenging Sonar task (distinguishing rocks from mines), the system achieved accuracies exceeding 97% (specifically $0.984 \pm 0.011$ for analogue RC). The highest performance was observed at the "coupling-edge" regimes where the system transitions between different numbers of self-pulsating rings.
Robustness and Reproducibility: These results were validated on a second SCISSOR device with slightly different fabrication parameters, confirming that the coupling-edge operating principle is reproducible across devices.
Performance vs. Complexity: The study found that regions of highest dynamical complexity (fully chaotic, high $\rho$ ) did not necessarily yield the best classification accuracy. Instead, intermediate complexity regions at the coupling edges provided the optimal balance of rich nonlinear dynamics and structured output.

Significance and Claims
The authors claim that optical coupled microring resonators, specifically in SCISSOR configurations, serve as effective building blocks for future edge computing and neuromorphic photonic systems. The work suggests that understanding the physical computing mechanisms of individual network nodes is essential for scaling. By identifying "coupling-edge" regimes as optimal operating points, the paper proposes a physics-driven criterion for selecting operating conditions. This approach significantly reduces the need for exhaustive parameter searches in future designs and offers a strategy for training and controlling scaled-up on-chip architectures. The ability to achieve high performance with low optical powers (< 8 mW for demanding regimes) and sparse digital representations positions this technology as a viable path toward efficient, real-time photonic inference.