[Paper Review] Silicon photonic-electronic neural network for fibre nonlinearity compensation

Chaoran Huang, Shinsuke Fujisawa+13 more|arXiv (Cornell University)|Oct 11, 2021

Neural Networks and Reservoir ComputingComputer Science54 references283 citations

TL;DR

This paper presents a reconfigurable silicon photonic-electronic neural network that performs real-time fiber nonlinearity compensation in a 10,080 km submarine optical transmission system. By implementing a photonic neural network with integrated Mach-Zehnder resonators and tunable photonic neurons on a CMOS-compatible platform, the system achieves Q-factor improvements comparable to software-based deep learning on a GPU, demonstrating sub-ns processing speed and low power consumption for high-bandwidth signal processing.

ABSTRACT

In optical communication systems, fibre nonlinearity is the major obstacle in increasing the transmission capacity. Typically, digital signal processing techniques and hardware are used to deal with optical communication signals, but increasing speed and computational complexity create challenges for such approaches. Highly parallel, ultrafast neural networks using photonic devices have the potential to ease the requirements placed on the digital signal processing circuits by processing the optical signals in the analogue domain. Here we report a silicon photonice-lectronic neural network for solving fibre nonlinearity compensation of submarine optical fibre transmission systems. Our approach uses a photonic neural network based on wavelength-division multiplexing built on a CMOS-compatible silicon photonic platform. We show that the platform can be used to compensate optical fibre nonlinearities and improve the signal quality (Q)-factor in a 10,080 km submarine fibre communication system. The Q-factor improvement is comparable to that of a software-based neural network implemented on a 32-bit graphic processing unit-assisted workstation. Our reconfigurable photonic-electronic integrated neural network promises to address pressing challenges in high-speed intelligent signal processing.

Motivation & Objective

• Address the challenge of fiber nonlinearity impairments in long-haul, high-capacity optical communication systems.
• Overcome the computational and power limitations of digital signal processing (DSP) for real-time, high-bandwidth optical signal processing.
• Develop a photonic neural network (PNN) platform that performs analog, real-time inference with gigahertz bandwidth and low latency.
• Demonstrate the feasibility of integrating full neural network functions—weighting, summation, and nonlinear activation—on a single silicon photonic chip.
• Achieve performance comparable to software-based neural networks on high-end GPUs, but with hardware advantages in speed and energy efficiency.

Proposed method

• Utilizes a CMOS-compatible silicon photonic platform with Mach-Zehnder resonators (MRRs) as tunable photonic weights for linear operations.
• Employs integrated photonic neurons with reconfigurable activation functions based on Lorentzian-shaped transfer functions from MRRs.
• Integrates on-chip photodetectors (germanium-on-silicon) and electronic components (resistors, capacitors) for full photonic-electronic integration.
• Employs N-doped photoconductive heaters to tune the resonance wavelengths of MRRs for programmable synaptic weights.
• Uses a hybrid architecture where pre-trained artificial neural network (ANN) parameters are uploaded to the photonic chip for inference.
• Employs wavelength-division multiplexing (WDM) to enable parallel processing of multiple signal channels on the same chip.

Experimental results

Research questions

• RQ1 Can a fully integrated photonic-electronic neural network perform real-time, high-bandwidth signal processing for fiber nonlinearity compensation?
• RQ2 How does the performance of a hardware-implemented photonic neural network compare to a software-based deep learning model on a GPU in terms of Q-factor improvement?
• RQ3 To what extent can photonic neurons with reconfigurable activation functions model complex nonlinear distortions in optical fiber links?
• RQ4 Can the photonic neural network maintain signal integrity and achieve low latency in a 10,080 km submarine transmission system?
• RQ5 What is the scalability of the photonic neural network in terms of neuron count and network depth for real-world signal processing tasks?

Key findings

• The photonic neural network achieved a Q-factor improvement in a 10,080 km submarine transmission system that is comparable to that of a software-based neural network running on a 32-bit GPU-assisted workstation.
• The system demonstrated real-time processing of optical signals in the analog domain, with processing delay limited only by the speed of light in the waveguides, enabling sub-ns operation.
• The use of Lorentzian-shaped transfer functions in photonic neurons enabled accurate modeling of nonlinear perturbations in fiber transmission, outperforming Leaky ReLU in simulation.
• The Q-factor improvement increased with the number of neurons in the second hidden layer, confirming the network's ability to model complex nonlinear distortions more precisely as capacity increases.
• The photonic-electronic integrated circuit achieved high performance with minimal added delay as the network size scaled, demonstrating scalability for real-time applications.
• The platform supports reconfiguration of activation functions and synaptic weights via electrical tuning, enabling programmability for different signal processing tasks.