Source-linked AI summary
Silicon photonic-electronic neural network for fibre nonlinearity compensation
Chaoran Huang, Shinsuke Fujisawa, Thomas Ferreira de Lima, Alexander N. Tait, Eric C. Blow, Yue Tian, Simon Bilodeau, Aashu Jha, F atih Yaman, Hsuan-Tung Peng, Hussam G. Batshon, Bhavin J. Shastri, Yoshihisa Inada, Ting Wang, Paul R. Prucnal
TL;DR
Conventional approaches face speed and computational-complexity requirements, while prior software neural networks were not real-time. The paper reports a reconfigurable silicon photonic-electronic neural-network platform that experimentally models fibre nonlinearity and achieves Q-factor improvement comparable to numerical simulations.
Problem
Conventional digital signal-processing circuits face demanding speed and computational-complexity requirements for high-speed signals.
Method
The paper uses a reconfigurable photonic-electronic integrated neural-network platform incorporating photonic components on silicon for neural-network computing.
Results
The extended system accurately models fibre nonlinearity, leading to Q-factor improvement comparable with numerical simulations on a conventional computer.
Takeaways & Limitations
The experimentally implemented system targets extremely high throughput and real-time signal processing for high-speed information processing.
Takeaways & Limitations
Prior artificial neural networks had only been implemented in software, so their data processing was not real-time.
Abstract
from arXiv · showhide
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.
munication system. The Q-factor improvement is comparable to that of a software-based
The paper presents a reconfigurable silicon photonic-electronic neural network for real-time, high-speed intelligent signal processing. It demonstrates fibre nonlinearity compensation in submarine communications while addressing conventional DSP complexity and speed constraints.
- Motivation: High-speed optical signals require real-time processing, but digital hardware demands massive parallelisation and creates circuit-complexity and power-budget constraints.These constraints have limited implementation of computationally intensive ANN algorithms.
- Motivation: Photonic neural networks offer highly parallel optical interconnects that can support real-time ANN implementations with tens of gigahertz bandwidth in a single pipeline.The approach targets DSP-related performance and computational-complexity constraints.
- Novelty: The work addresses a gap in prior photonic neural networks, where linear operations and nonlinear activation functions were not fully integrated for gigahertz-bandwidth real-time processing.The paper identifies integrated fast photonic neurons as a critical step toward real-time applications such as fibre nonlinearity compensation.
- System: The integrated silicon PNN combines neurons and synapses to provide accurate weighting, summation, and biased nonlinearity for neural-network computing.The system is reconfigurable and integrates a full neural-network model on a silicon photonic platform.
- Application: The platform was applied to fibre nonlinearity compensation in submarine optical-fibre communication systems, with ANN parameters trained on a computer and uploaded to the photonic chip for inference acceleration.This establishes a photonic implementation of the compensation task rather than a software-only ANN.
- Results: The system improved signal quality in a 10,080 km submarine fibre communication system, with Q-factor improvement comparable to a software neural network on a 32-bit GPU-assisted workstation.The results indicate preserved signal integrity while offering potential speed and gigahertz-bandwidth advantages.
1 Neural network model for fiber nonlinearity compensation
The paper models fibre nonlinearity as a nonlinear perturbation estimated from symbol triplets and removed from received symbols using a feed-forward neural network. The model targets high-throughput compensation for long-distance optical transmission, where nonlinear impairments challenge digital hardware.
- 10,080 km submarine transmission is studied with a single-channel 32 Gbaud polarization-multiplexed 16-QAM signal.
- Fibre nonlinearities remain a major limiting impairment because computationally expensive nonlinear compensation is difficult to implement with ASICs.
- The nonlinear impairment is represented as a perturbation of the linearly propagated optical field and estimated from transmission data.
- The NN uses nonlinear triplets as inputs and predicts the real and imaginary components of the estimated nonlinearity for each symbol.
- The recovered symbol is obtained by subtracting the estimated nonlinear perturbation from the received symbol, with training minimizing mean squared error against the transmitted symbol.
- Processing 32 Gbaud PM-16QAM signals requires approximately 232.4 TOPS, exceeding the cited 92-TOPS peak throughput of a typical TPU.
2 Photonic hardware implementation
The neural network is implemented with a wavelength-division-multiplexed silicon photonic architecture that performs weighted broadcast, summation, and nonlinear activation on chip. The demonstrated photonic circuit implements the second hidden layer and supports configurable weights and biases.
- The PNN uses a WDM-based broadcast-and-weight architecture on a standard silicon photonic platform.
- The experimentally implemented 4×2 PNN realizes the NN’s second hidden layer and demonstrates fast weighting, summation, and biased nonlinearity.
- The authors state that the PNN can be robustly scaled to incorporate the full NN model.
- Two MRR weight-bank arrays connect photonic neurons, with each second-layer neuron receiving a 2×2 weighted combination from the first layer.
- MRR transmission is thermally tuned with N-doped heaters to configure connection strengths, providing continuous multichannel control with up to 8-bit accuracy.
- Balanced photodetection subtracts complementary MRR outputs to provide a continuous -1 to +1 weight range.
- The MRR modulator’s nonlinear electrical-to-optical transfer function supplies the neuron activation, while electrical bias controls its operating point.
3 Training with photonic neural network
Training characterizes the photonic neuron activation function before optimizing the neural network with that device response. The trained model is evaluated across launch powers on a 10,080 km fibre link.
- The photonic activation function is measured from on-chip neuron waveforms, fitted to a Lorentzian function, and included during Adam-based training.
- 32,106 training symbols from a 32 Gbaud PM-16-QAM signal are transmitted over a 10,080 km pure-silica-core fibre link.
- The weights are constrained to -1 to 1 to match the operating range of the MRR weight bank.
- The independent cross-validation Q-factor gradually increases during training and converges to 8.1 dB after 21,600 steps.
- Testing uses varying launch powers, with the estimated nonlinear perturbation scaled by the training-to-test launch-power difference before subtraction.
- The simulated PNN increases Q-factor by 0.66 dB without NLC and achieves a Q-factor comparable to 3-step digital backpropagation.
4 Photonic neural network implementation
The trained network is mapped onto the photonic chip and tested against a simulated neural network. The physical PNN compensates fibre nonlinearity while retaining nearly the simulated model’s Q-factor gain.
- The measured photonic neuron bandwidth is 150 MHz, with simulations indicating practical improvement to 10 GHz using optimized modulator geometry.
- The trained parameters are mapped as four 2×2 current matrices and two neuron biases applied simultaneously to the PNN chip.
- Photonic-neuron outputs are compared with a 32-bit GPU-assisted simulated NN using mean squared error over 32,016 test symbols.
- 0.008 to 0.014 is the reported range of photonic-neuron mean squared errors caused by noise and deviations from target weights and activation functions.
- 0.60 dB is the measured PNN Q-factor improvement, compared with a 0.65 dB gain from the simulated NN.
- 0.05 dB is the reported penalty for loading the neural network onto the PNN after accounting for physical-component and equipment noise.
5 Scalability for NLC
The paper discusses scaling the photonic neural network for fibre nonlinearity compensation through footprint reduction, integration, calibration, and suitable light sources. Device non-uniformity and electrical interfacing remain important scalability considerations.
- The photonic neural network can be implemented using standard silicon photonic platforms and integrated optoelectronic components on one chip.
- Weight pruning reduces the NN model from 892 input triplets to 300, while the resulting network contains 952 optoelectronic components.
- The 952 optoelectronic components require 0.38 mm2 of chip space, with electrical traces and pads adding packaging-dependent area.
- Electrical pad area can reach 0.76 mm2, with each bonding pad occupying 20µm × 20µm; comparable silicon photonic footprints have been demonstrated.
- Device non-uniformity changes the relationship between MRR currents, actual weights, and activation-function shapes across the chip.
- Calibration, trimming, in-situ training, frequency-comb WDM sources, and soliton-based CMOS-compatible circuits are proposed to address scaling challenges.
6 Claiming cascadability for NLC
The paper frames cascadability as the ability of one photonic neuron layer to drive the next while limiting noise propagation. It identifies sufficient power, optical or electrical gain, and nonlinear noise suppression as requirements.
- Cascadability means that one neuron output can excite neurons in the subsequent layer.
- Cascadability requires equivalent responses in the subsequent layer and avoidance of noise propagation across multiple neural layers.
- Enough power must compensate for loss and power splitting caused by fan-out when one layer drives the next.
- The required power can come from optical lasers or electrical gain supplied by a high-transimpedance amplifier, passive resistor, or active TIA.
- A sufficiently large modulator voltage swing enables the nonlinear transfer function to suppress noise and avoid propagation through the network.
7 Impact of activation function and the number of neurons
Activation-function choice affects fibre nonlinearity compensation because the perturbation produces heavy-tailed symbols. The paper also reports improved Q-factor with more second-hidden-layer neurons and little added delay as network size grows.
- Fibre nonlinearity perturbation produces long, heavy-tailed symbols associated with nonlinear phase noise.
- An activation function with an unbounded range is preferred for extracting tail-edge features in the received symbols.
- Leaky ReLU has no upper boundary for positive input, creating a performance gap relative to the Lorentzian activation of an MRR neuron.
- Activation functions must be selected to synthesize particular neural-network tasks, while multiple coupled cavities could provide transfer functions beyond the Lorentzian shape.
- Q-factor improves as the number of neurons in the second hidden layer increases because the network can represent more parts of the nonlinearity perturbation.
- Photonic neural networks add almost no delay as NN size increases because their major operations are computed at a single time step.
8 Conclusion
The study presents a reconfigurable silicon photonic-electronic neural network that performs core neural computations and compensates fibre nonlinearities in submarine links. Its simulated extended system achieves Q-factor improvement comparable to conventional computer-based neural-network simulations while targeting high-speed analogue processing.
- The reconfigurable integrated platform could support a range of high-speed intelligent signal-processing problems.
- The platform integrates photonic components on silicon for high-speed information processing.
- It performs essential neural-network operations, including accurate weighting, summation, and biased nonlinearity.
- The system was applied to fibre nonlinearity compensation in trans-Pacific transmission links requiring extremely high throughput and real-time processing.
- The extended neural network accurately models fibre nonlinearity, leading to Q-factor improvement comparable with numerical simulations on a conventional computer.
- Analogue optical-signal processing can relax the complexity and speed requirements typically imposed on conventional DSP circuits.
Methods
The device is fabricated on a silicon-on-insulator platform with microring-resonator weight banks, modulators, heaters, photodetectors, and electrical interconnects. The methods specify the component geometries, doping structures, and passive electrical values used in the integrated system.
- The device is fabricated on a silicon-on-insulator wafer with 220 nm silicon and 2 µm buried oxide thicknesses.
- The weight bank uses four microring resonators in an add/drop configuration, with radii of 8.0, 8.1, 8.2, and 8.3 µm.
- Only the first two microring resonators are used experimentally, while differing radii avoid resonance collision.
- The ring-to-bus-waveguide gap is 200 nm, yielding a Q factor of ∼6000, and in-ring N-doped photoconductive heaters control the weights.
- The microring modulator uses an 8 µm-radius ring coupled to two bus waveguides with 0.2 and 0.5 µm gaps.
- Germanium-on-silicon photodetectors use a deposited germanium layer with implanted boron and phosphorus forming horizontal p-i-n junctions and ohmic contacts.