Real-time all-optical signal equalisation with silicon photonic recurrent neural networks

Ruben Van Assche, Sarah Masaad, Emmanuel Gooskens, Stijn Sackesyn, Joris Van Kerrebrouck, Xin Yin, Peter Bienstman · 2025

Read Real-time all-optical signal equalisation with silicon photonic recurrent neural networks by Ruben Van Assche, Sarah Masaad, Emmanuel Gooskens, Stijn Sackesyn, Joris Van Kerrebrouck, Xin Yin, Peter Bienstman (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-19911.txt.

Introduction

Ruben Van Assche and six co-authors clean up distorted light signals with a chip that computes with light itself. Their paper appeared in two thousand twenty-five. Signals in optical fibres blur and warp from dispersion and nonlinear effects. Digital cleanup works, but it burns power and adds delay. The team pursues an all-optical fix instead.

Their silicon photonic chip joins a recurrent reservoir with a tunable readout layer. It is, to their knowledge, the first live online demo of its kind. The chip restored a fast on-off keying signal across varied power levels and fibre lengths. It coped even where the fibre behaved in a strongly nonlinear way. The result points to lower-power, lower-delay links for future networks.

Catalog record

CREATORRuben Van Assche, Sarah Masaad, Emmanuel Gooskens, Stijn Sackesyn, Joris Van Kerrebrouck, Xin Yin, Peter Bienstman
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-19911

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Real-time all-optical signal equalisation with silicon photonic recurrent neural networks Ruben Van Assche, Sarah Masaad, Emmanuel Gooskens, Stijn Sackesyn, Joris Van Kerrebrouck, Xin Yin, Peter Bienstman Abstract Communication through optical fibres experiences limitations due to chromatic dispersion and nonlinear Kerr effects that degrade the signal. Mitigating these impairments is typically done using complex digital signal processing algorithms. However, these equalisation methods require significant power consumption and introduce high latencies. Photonic reservoir computing (a subfield of neural networks) offers an alternative solution, processing signals in the analog optical domain. In this work, we present to our knowledge the very first experimental demonstration of real-time online equalisation of fibre distortions using a silicon photonics chip that combines the recurrent reservoir and the programmable readout layer. We successfully equalize a 28 Gbps on-off keying signal across varying power levels and fibre lengths, even in the highly nonlinear regime. We obtain bit error rates orders of magnitude below previously reported optical equalisation methods, reaching as low as 4 × 10 − 7 4\text{\times}{10}^{-7} , far below the generic forward error correction limit of 5.8 × 10 − 5 5.8\text{\times}{10}^{-5} used in commercial Ethernet interfaces. Also, simulations show that simply by removing delay lines, the system becomes compatible with line rates of 896 Gpbs. Using wavelength multiplexing, this can result in a throughput in excess of 89.6 Tbps. Finally, incorporating non-volatile phase shifters, the power consumption can be less than 6 fJ/bit. 1 Introduction Figure 1 : Silicon photonics integrated reservoir (a) Photograph of the photonic chip on a printed circuit board with a one-euro coin next to it for scale. (b) Diagram of signal equalisation using reservoir computer (PIC: Photonic Integrated Circuit). Global data traffic has increased exponentially over the past twenty years [ 1 ] . Although optical networks were initially primarily designed for long-distance, high-speed transmission, emergent novel applications such as the Internet of Things, cloud services, artificial intelligence, and edge computing have drastically increased the amount of short-length inter- and intra-data center traffic. For long-reach optical networks, wavelength division multiplexing and coherent modulation formats are used frequently and can achieve high data rates and excellent spectral efficiency. Examples of these formats are “polarization division multiplexing quadrature phase shift keying” (PDM-QPSK) and “polarization division multiplexing quadrature amplitude modulation” (PDM-QAM). However, the transceivers for these formats are costly, which makes them unsuitable for short-haul networks, where the rapid increase in demand for network capacity has made deployment and operational costs a critical factor [ 2 ] . Currently, “intensity modulation with direct detection” (IM/DD) is the transceiver technology used for short-reach networks. However, this limits the achievable bit rate, since IM/DD only encodes data on the intensity of the signal, compared to coherent communication where information is encoded on the amplitude, phase, and/or polarization of the signal. To meet the increasing demands of the industry, it is essential to develop solutions that enhance data rates and power efficiency of IM/DD systems, while avoiding more costly coherent modulation formats. When considering fibre-optic communication systems, a large constraint for the throughput of IM/DD systems stems from the distortions arising from propagation through optical fibres. These distortions can be subdivided into linear distortions, due to chromatic dispersion and polarization-mode dispersion, and nonlinear distortions due to the Kerr effect. These Kerr nonlinearities in particular pose challenges because of the self-phase modulation they induce. In a single channel, it is the interplay between chromatic dispersion and self-phase modulation that severely degrades the quality of the signal coming out of the fibre. State-of-the-art approaches utilize digital signal processing (DSP) techniques to mitigate these issues. Although effective, these algorithms introduce additional communication latency, as they require processing in the electrical domain. Moreover, their computational complexity leads to significant power consumption, presenting challenges for energy efficiency and operational costs in short-haul optical networks. Their performance for signals with high nonlinear distortions is limited as well [ 3 , 4 ] . Extensive research has been conducted to explore alternative methods for addressing these challenges. For compensating linear dispersion, solutions have been developed both in the electric and optical domains, many of which are now widely adopted. Examples include dispersion compensated/shifted fibres [ 5 ] and fibre Bragg gratings [ 6 ] in the optical domain, and fractionally spaced equalizers [ 7 ] and feed-forward equalizers [ 8 ] in the digital electronic domain. For the nonlinear part of the problem, electrical and optical solutions have been found as well. In the electric domain, digital backpropagation [ 9 ] or nonlinear Volterra series [ 10 ] are notable solutions. These implementations are power hungry however, due to the complexity of their algorithms, and introduce a latency due to their digital implementation. Optically the most important examples are optical phase conjugation (OPC) [ 11 ] and phase-conjugated twin waves (PC-TW) [ 12 ] . However, OPC relies on an a priori known fibre length, which means it cannot be used for an all-purpose solution, and PC-TW sacrifices half the transmission capacity. In addition to these implementations, research has been conducted to try and solve these problems using machine learning and neural networks, both in the digital electronic domain, as well as in the optical domain [ 13 , 14 , 15 , 16 , 17 , 18 , 19 ] . An intuitive approach for machine learning in this context is to implement a neural network after the optical fibre or after the receiver and to consider this neural network as a post-processor that retrieves the original undistorted signal. One promising candidate for an optical implementation of this concept is reservoir computing (RC). This is a machine learning paradigm that uses a fixed, randomly initialized recurrent neural network, followed by a simple trainable linear readout. In practice, the random initialization typically stems from fabrication imperfections. Moreover, the simple trainable linear readout has the advantage of being easy to train, as well as keeping power consumption limited. Within the reservoir paradigm, there are two possible implementations: time-delay-based reservoirs (TDRC) and spatially distributed reservoirs [ 20 ] . Experiments and simulations utilizing TDRCs have already shown the ability for optically equalizing dispersion [ 21 , 22 , 23 , 24 , 25 ] . However, these systems require time multiplexing, which introduces challenges at high speeds due to the response time of the nonlinear neuron. This means that for TDRCs a trade-off exists between the computational capacity and the processing rate [ 26 ] [ 27 ] [ 28 ] . This issue is not present for spatially distributed reservoirs, where we showed previously that a linear and passive integrated circuit as a reservoir, in combination with a separate digital readout layer, forms a suitable solution for signal equalisation [ 29 ] . Additionally, in terms of bandwidth, a single chip with a single set of weights can simultaneously equalize signals at multiple wavelengths [ 30 ] . However, both of these demonstrations did not have functional optical weights and all output channels needed to be measured sequentially and processed digitally offline. As a result, these systems did not operate in real time, despite the high-speed optical computation in the reservoir itself. There was an initial attempt to integrate a readout layer, which however was

Read the complete text