Experimental Demonstration of Zero-Shot Machine Learning Equalisation in Dual-Polarisation Coherent Transmission

Aug 4, 2026

Upload by Sage

Introduces the first machine-learning-only equaliser for dual-polarisation IQ-modulated signals that performs without any online training or fine-tuning. Laboratory transmission at 30 Gbaud with DP-16-QAM demonstrates equaliser performance that can match or exceed conventional DSP across diverse conditions, including different fibre types and transmission frequencies. The approach targets adaptability in optical fibre links by using static parameters rather than continual channel-dependent re-training.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Abstract

Introduction

Machine learning equalisation criteria

Limitations of existing adaptive approaches

Transfer learning and pre-training

Motivation for zero-shot experimental results

Fibre Channel Modelling & Training Sequences

Linear simulations with Jones transfer matrix

Inference inputs and known training sequence

Pilot design and modulation/filter parameters

Experimental Demonstration of Zero-Shot Machine Learning Equalisation in Dual-Polarisation Coherent Transmission Samuel Lennard, Fabio A. Barbosa, Filipe M. Ferreira Optical Networks Group, Dept. Electronic & Electrical Eng. , University College London (UCL), UK, [s.lennard@ucl.ac.uk](s.lennard@ucl.ac.uk) Abstract We introduce the first machine learning-only equaliser for dual-polarisation IQ-modulated signals operating without online training or fine-tuning. Lab transmission of 30 Gbaud DP-16-QAM shows this equaliser matching or outperforming conventional DSP over a range of conditions, including different fibres and transmission frequencies. ©2024 The Author(s) Introduction Machine learning (ML) techniques have recently risen to prominence as a potential route for fibre channel equalisation. In order for an ML equaliser to be considered effective, it must meet three criteria: it must be highly performant, it must be highly adaptable, and it must have low complexity[1][2] . There has been extensive work addressing the first point, showcasing neural network (NN)-based equalisation achieving significant gains over conventional digital signal processing (DSP) methods, particularly within the nonlinear regime[3] . In addition, recent works have started to consider the issue of complexity within NNs, attempting to lower complexity to an acceptable level[4] . Despite the performance demonstrated by ML methods, the problem of adaptability is yet to be solved. While these NNs have been effective, they often require a significant amount of training data in order to adapt to a channel. In addition, complexity calculations often neglect to take into account the added overhead of performing backpropagation and gradient descent, which would need tobe carried out regularly for an evolving channel[5] . This prevents these methods from being applied in a dynamic network setting or with particularly turbulent channels. In addition, this also limits NNs from being applied to equalizing stochastic polarisation and laser effects, as adapting to these would require frequent re-training. One way to increase the adaptability of ML equalisers is to minimise the amount of training that needs to be carried out. This can be achieved through pre-training the NN in simulation before being applied to a real fibre channel, allowing for the bulk of the training to be carried out without any online training and data acquisition. The work in[6] demonstrates the use of transfer learning from simulation to initialise NN parameters before training on experimental data for nonlinear equalisation, allowing for strong performance on multi-channel transmission with minimal online training. In a similar vein,[1] and[2] also show transfer learning being used to reduce the necessary training set size by up to 99% . Lastly,[7] fully mitigated online learning by pre-training in simulation, and showcased that the NN was able to generalise over the range of parameters it was trained on and carried out successful nonlinearity mitigation, also in simulation. However, while these papers showcase that generalisation is possible, there still lacks any work that demonstrates zero online training whilst maintaining high performance in experimental settings. In addition, these papers focus on equalising deterministic effects such as nonlinearities and chromatic dispersion (CD), neglecting stochastic effects such as polarisation mode dispersion (PMD) and laser phase noise, where further performance and complexity improvements may be acquired. This paper addresses these issues by constructing an NN that does not require any online training and instead possesses static parameters that are capable of equalizing over a range of channel conditions, including stochastic effects, and therefore maximizing adaptability. We demonstrate the NN’s ability to perform in an experimental, zero-shot setting-that is, having never been trained on any experimental data prior to

FAQs

What makes the proposed equaliser “zero-shot” in experimental transmission?

It uses static network parameters and performs inference without any online training or fine-tuning, meaning it has not been trained on the experimental data before running inference.

How is the training performed prior to inference?

A comprehensive training phase is carried out using linear transmission simulations (Jones transfer matrix) with sampled channel and stochastic-process parameters, rather than adapting during experiments.

What transmission conditions are tested to show adaptability?

Results are reported across multiple conditions, including different fibre types, transmission frequencies, and polarisation delay, while keeping the same neural network.

Aug 4, 2026

Upload by Sage