A One-Dimensional Convolutional Neural Network-Based Equalizer for 100 Gb/s Short-Reach Optical Communication Systems

Abstract

Short-reach optical communication systems that use intensity modulation and direct detection (IM/DD) with PAM-4 signaling are considered as a desirable option due to its simple architecture and lower cost. However, bandwidth limitations, chromatic dispersion and device nonlinearities cause serious signal distortions, which degrade overall system performance. In this paper, we present a one-dimensional convolutional neural network (1D-CNN)-based equalizer to address the challenges of traditional feed-forward equalizers (FFE), which are fundamentally limited to linear impairment compensation. The proposed equalizer is designed with a lightweight CNN architecture that maintains structural simplicity while effectively exploiting temporal features from the received symbol sequences. Simulation results demonstrate that the 1D-CNN equalizer can improve receiver sensitivity by nearly 2.5 dB compared to the conventional FFE and by around 1 dB compared to the artificial neural network (ANN)-based equalizer, at the same BER level. Furthermore, we investigate the impact of network depth and feature map size on equalization performance, providing practical insights for real-world deployment.