Source-linked AI summary
Photonic machine learning implementation for signal recovery in optical communications
Apostolos Argyris, Julián Bueno, Ingo Fischer
TL;DR
Time-dependent, high-speed signals become difficult to process when nonlinearly distorted, motivating signal-recovery approaches beyond standard tools. The paper investigates ELM and reservoir computing for recovering distorted signals, with an experimental implementation extending transmission to 51 km, while remaining short of established signal-processing methodologies.
Problem
Processing time-dependent, high-speed signals is challenging when signals are nonlinearly distorted, and standard tools have drawbacks in ultrafast systems.
Method
The paper applies techniques including extreme learning machines and reservoir computing to data recovery of distorted signals.
Results
51 km transmission length is achieved with the experimental implementation.
Takeaways & Limitations
ELM and reservoir computing offer solutions for data recovery of distorted signals, although the demonstrated performance remains below established signal-processing methodologies.
Takeaways & Limitations
The approach does not yet reach the status of well-established methodologies in signal processing, and standard tools have drawbacks in ultrafast systems.
Abstract
from arXiv · showhide
Machine learning techniques have proven very efficient in assorted classification tasks. Nevertheless, processing time-dependent high-speed signals can turn into an extremely challenging task, especially when these signals have been nonlinearly distorted. Recently, analogue hardware concepts using nonlinear transient responses have been gaining significant interest for fast information processing. Here, we introduce a simplified photonic reservoir computing scheme for data classification of severely distorted optical communication signals after extended fibre transmission. To this end, we convert the direct bit detection process into a pattern recognition problem. Using an experimental implementation of our photonic reservoir computer, we demonstrate an improvement in bit-error-rate by two orders of magnitude, compared to directly classifying the transmitted signal. This improvement corresponds to an extension of the communication range by over 75%. While we do not yet reach full real-time post-processing at telecom rates, we discuss how future designs might close the gap.
1 Instituto de Física Interdisciplinar y Sistemas Complejos IFISC (CSIC-UIB), Campus UIB, 07122, Palma de
The paper addresses recovery of high-speed optical signals distorted by fibre transmission using a simplified photonic reservoir that reframes bit detection as pattern recognition. Experimentally, this approach substantially reduces BER and extends usable transmission distance, while remaining short of full real-time telecom-rate processing.
- Severely distorted, time-dependent optical signals challenge conventional high-speed processing, particularly under nonlinear fibre impairments.
- The method remains constrained by computationally expensive standard processing and does not yet match established signal-processing methodologies or full telecom-rate real-time operation.
- The proposed simplified reservoir converts direct bit detection into pattern recognition using nonlinear transient responses and sequential photonic processing.
- BER is reduced to 10-2 when neighbouring bits inform pattern recognition, improving on direct classification of the transmission output.
- 1.8∙10-4 BER is achieved with the experimentally implemented photonic reservoir, while an ELM configuration reaches 7∙10-4.
- 75.9% transmission-distance gain over classifying the transmission output and 200% over direct detection are obtained in the short-reach system.
- The approach extends short-reach transmission from 17km with direct detection to 29km with linear regression and 51km with the photonic reservoir.
Methods
The study models distorted fibre-optic transmission and processes the detected signal with a simplified photonic reservoir whose transient responses feed a linear classifier. Experimental and numerical procedures vary masking, virtual-node dimensions, training windows, and operating conditions to recover transmitted bits.
- Transmission and signal preparation: Fibre transmission is modelled with a coupled nonlinear Schrödinger equation, including attenuation, chromatic dispersion, Kerr nonlinearity, amplification, detection, and electrical filtering.The numerically generated signals feed both the experimentally built reservoir and numerical RC investigations.
- Input encoding and reservoir representation: Each transmitted bit is represented by an analogue pattern of j samples, multiplied by a random mask before entering the reservoir.When j=k, the input and reservoir state dimensions match; choosing j<k increases the state-space dimensionality when k is a multiple of j.
- Mask selection and limitations: For short reservoirs, mask choice becomes critical, especially when k≤16, so BER maps average or select among ten random masks.Experimental scenarios showed insignificant dependence on the tried random mask sequences, whereas numerical maps reported the masks giving the lowest BER.
- Classifier training: The reservoir is divided into virtual nodes, and node responses are used to train a ridge-regression linear classifier on bit streams separated into training, validation, and independent test sets.The procedure uses 75% of one stream for training, 25% for cross-validation, 10 repetitions, and a ridge parameter of 0.01.
- Temporal readout: Predictions can combine responses from neighbouring bit timeframes because fibre nonlinearities spread information across adjacent bits.The number of previous and consecutive responses is selected according to the distortion and introduces latency for future-bit responses.
Materials & Correspondence
The paper states that supporting data are available from the corresponding author on reasonable request, while materials requests should be directed to the named authors.
- Study data are available from the corresponding author on reasonable request.
- Requests for materials should be addressed to A.A. and I.F.
Supplementary Figures
The supplementary material describes short-reach and long-haul transmission systems, signal distortions, reservoir operating conditions, BER mappings, and comparisons with direct detection. It reports substantial BER improvement under selected reservoir configurations and identifies operating regimes and transmission boundaries that constrain performance.
- Short-reach system: The short-reach system uses 25Gb/s random PAM-NRZ transmission over standard single-mode fibre with PIN detection and Butterworth electrical filtering.
- Long-haul system: The long-haul system uses 40 serial 100km modules with dispersion-compensating fibre, EDFAs, and optical filtering, reaching z2=4000km of SSMF.
- Transmission limits: Without equalization or post-processing, decoded BER exceeds 0.1 beyond z1>41km for short-reach and z2>3300km for long-haul transmission.
- Signal distortion: At z1=45km and z2=4000km, chromatic dispersion, Kerr nonlinearity, and amplification noise eliminate efficient binary-level separability.
- Operating boundaries: Complete injection locking and sufficiently large detuning produce poor BER improvement, while long-haul operation is optimized at Δf=0GHz, 16dB excess attenuation, and a 4τ response.
- Short-reach operating conditions: Short-reach BER improvement is significant under partial locking and moderate feedback, requiring at least 5τ of reservoir response and maximizing at 9τ with BER=1.8∙10^-4.
- BER comparison: Direct detection without dispersion compensation has BER>0.2 across the investigated optical-SNR range, whereas reservoir processing yields BER<10^-4 under the reported optimized conditions.