All-Optical Reservoir Computing on a Photonic Chip Using Silicon-Based Ring Resonators

Florian Denis-le Coarer, Marc Sciamanna, Andrew Katumba, Matthias Freiberger, Joni Dambre, Peter Bienstman, Damien RontaniIEEE Journal of Selected Topics in Quantum Electronics 2018 · DOI

TL;DR

This paper presents a novel photonic chip architecture for reservoir computing, utilizing silicon microring resonators as nonlinear nodes. The proposed system demonstrates high performance on a challenging nonlinear task (delayed XOR) at 20 Gb/s with low bit error rates and remarkably low power consumption, paving the way for energy-efficient, high-speed optical computing.

What Problem Does This Paper Address?

The ever-increasing volume of data generated by modern applications necessitates faster and more energy-efficient computing solutions. Traditional electronic hardware faces limitations in speed and power consumption. Reservoir computing, a machine learning paradigm, offers a promising approach for hardware implementation due to its simplified training process. Photonics, with its inherent speed and low loss, is an attractive medium for building reservoir computers. However, creating effective photonic reservoirs that are both high-performing and energy-efficient remains a challenge. This paper addresses this by proposing and analyzing an all-optical reservoir computing architecture on a silicon chip, specifically using nonlinear microring resonators as the core computational elements.

Methods

The researchers developed a 4x4 "swirl-topology" photonic reservoir. The key innovation lies in using silicon microring (MR) resonators as the nonlinear nodes. These MRs are integrated onto a silicon chip, allowing for compact and scalable designs. The reservoir's state update is governed by a set of nonlinear dynamical equations that model the behavior of these MRs, taking into account effects like two-photon absorption, free carrier dispersion, and losses.

The performance of this reservoir was evaluated on the delayed XOR task, a standard benchmark for nonlinear processing. They conducted extensive numerical simulations to analyze the impact of various design parameters:

  • Interconnection lengths (waveguide delays): How the time it takes for signals to travel between nodes affects performance.
  • Injection parameters:
    • Injected power: The optical power fed into each microring resonator.
  • Optical detuning: The difference between the injected light's frequency and the microring's resonance frequency.

They also compared their "MR-reservoir" to a previously proposed "passive reservoir" that used linear optical components.

Datasets

The paper uses a synthetic bit stream for the delayed XOR task, consisting of 20,000 random bits. This stream is divided into training (16,000 bits) and testing (4,000 bits) sets. The training is performed using regularized ridge regression. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Nonlinear nodes are crucial for performance: Using nonlinear microring resonators as nodes, rather than passive linear components, significantly improves the reservoir's ability to handle complex nonlinear tasks like the delayed XOR.
  • Operating point matters: The performance of the reservoir is highly sensitive to the operating point of the individual microring resonators. They found that operating the MRs close to a stability bifurcation point is optimal.
  • Optical detuning is a powerful control knob: Tuning the frequency of the injected light away from the microring's resonance frequency (optical detuning) is critical for achieving good performance. Detuning, particularly negative detuning, can lead to better results than injecting light directly at resonance.
  • Robustness to imperfections: The proposed MR-reservoir shows good robustness to variations in the resonance frequencies of the individual microring resonators, a crucial factor for practical integrated photonic devices.
  • Energy efficiency is achievable: The architecture achieves high performance with remarkably low average injection power, significantly reducing the energy footprint compared to previous approaches.

Key Results

  • High-speed operation: The MR-reservoir successfully performs the delayed XOR task at 20 Gb/s.
  • Low bit error rates (BER): Achieved BERs are consistently below 10⁻³ for a wide range of parameters.
  • Low power consumption: The average injection power required is less than 2.5 mW (0.15 mW per node).
  • Comparison to passive reservoirs: The MR-reservoir achieves comparable or better performance than passive reservoirs, but with the added benefit of integrated nonlinearity.
  • Optimal detuning: Performance is generally better when the optical detuning is non-zero, with negative detuning showing a slight advantage.
  • Inter-delay optimization: The optimal interconnection length (inter-delay) is found to be around 18.75 ps, which is close to the internal time scale of the microring resonator's optical mode.

Limitations and Future Work

  • Limited task scope: The study focuses on a single nonlinear Boolean task (delayed XOR). The performance on other complex tasks like time series generation or chaos prediction needs further investigation.
  • Idealized simulations: While some simulations considered heterogeneities in resonance frequencies, further work could explore other imperfections and their impact.
  • Potential for improvement: The authors suggest that better management of losses within the integrated architecture could further enhance performance.
  • Future research directions: The paper opens avenues for exploring this type of integrated reservoir for other applications such as time series generation, chaos prediction, or nonlinear channel equalization. They also propose reducing the number of injected nodes for even greater energy efficiency.

Summary updated

Abstract

We present in our work numerical results on the performance of a 4 × 4 swirl-topology photonic reservoir integrated on a silicon chip. Nonlinear microring resonators are used as nodes. We analyze the performance of such a reservoir on a classical nonlinear Boolean task (the delayed XOR task) for: various designs of the reservoir in terms of lengths of the waveguides between consecutive nodes, and various injection parameters (injected power and optical detuning). From this analysis, we find that this kind of reservoir can perform-for a large variety of parameters-the delayed XOR task at 20 Gb/s with bit error rates lower than 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> and an averaged injection power lower than 2.5 mW.
BibTeX
@article{coarer2018all,
  author = {Florian Denis-le Coarer and Marc Sciamanna and Andrew Katumba and Matthias Freiberger and Joni Dambre and Peter Bienstman and Damien Rontani},
  title = {{All-Optical Reservoir Computing on a Photonic Chip Using Silicon-Based Ring Resonators}},
  journal = {IEEE Journal of Selected Topics in Quantum Electronics},
  year = {2018},
  doi = {10.1109/jstqe.2018.2836985},
}
Figure