Performance and Complexity Tradeoffs of Neural Network Equalizers in Intensity-Modulation Direct-Detection Optical Systems
Paper in proceeding, 2026
Increasing data center traffic has driven intensity-modulation direct-detection (IM/DD) systems toward symbol rates beyond 200 Gbaud, where bandwidth limitations and nonlinear impairments significantly degrade performance. Conventional equalizers often reach their practical limits under such conditions. Neural network (NN) equalization has therefore been widely investigated for impairment compensation in high-speed IM/DD links. However, improved signal recovery often comes at the cost of increased computational complexity, limiting real-time deployment. This work reviews experimental demonstrations of NN equalization in IM/DD systems and compares models under consistent link and training conditions. Bit-error-rate (BER) and multiplications per symbol (MPS) are evaluated jointly to quantify practical performance-complexity tradeoffs. The paper provides a reference for designing NN equalizers for IM/DD systems under receiver constraints.
intensity-modulation direct-detection
computational complexity
Neural network equalization
optical interconnects
Author
Dan Li
RISE Research Institutes of Sweden
Royal Institute of Technology (KTH)
Yevhenii Osadchuk
University of Copenhagen
Carlos Natalino Da Silva
Chalmers, Electrical Engineering, Communication, Antennas and Optical Networks
Armands Ostrovskis
Keysight Technologies Deutschland GmbH
Riga Technical University
Tianyu Jiang
Royal Institute of Technology (KTH)
RISE Research Institutes of Sweden
Darko Zibar
Technical University of Denmark (DTU)
Lu Zhang
Zhejiang University
Xianbin Yu
Zhejiang University
Vjaceslavs Bobrovs
Riga Technical University
Xiaodan Pang
Zhejiang University
Francesco Da Ros
Technical University of Denmark (DTU)
Oskars Ozolins
Riga Technical University
MACHINE LEARNING IN PHOTONICS II
0277-786X (ISSN) 1996-756X (eISSN)
Vol. 14104 1410409979-8-9023-2167-5 (ISBN)
2026 Conference on Machine Learning in Photonics
Strasbourg, France,
Strasbourg, France,
Hollow core fiber (HFC) for 6G mobile networks
VINNOVA (2024-02451), 2024-12-01 -- 2027-08-21.
Photonic-Assisted Hardware for Reservoir Computing (BRAIN)
Swedish Research Council (VR) (2022-04798), 2023-01-01 -- 2026-12-31.
Subject Categories (SSIF 2025)
Telecommunications
Signal Processing
DOI
10.1117/12.3097486