Performance Comparison of Waveform-, Symbol- and Bit-Levels of End-to-End Deep Learning Based Optimization for IM/DD Optical Interconnection

Deep-learning based end-to-end optimization provides a new design paradigm for optical interconnection systems with high communication capacity requirements. By jointly training the neural networks at the transceiver, end-to-end scheme has the advantage of finding globally optimal configurations. In this work, we construct a comparison framework with physical interpretability to evaluate different end-to-end schemes proposed in recent years. Based on neural networks' dimension, we classify end-to-end schemes into waveform-, symbol- and bit level respectively. We first optimize the loss function and training process to address the common defects found in previous work using bit- and symbol-level end-to-end schemes. Then, we conduct a comprehensive comparison among these end-to-end schemes and with the baseline using post-equalizer only. The impact of different modulation formats and modulation orders on system performance are also discussed. Finally, we explore the performance boundary of end-to-end schemes under optimal modulation configuration. From the experiments conducted over 0.5-km single-mode fiber, the bit-level end-to-end scheme achieves a highest bitrate of 336 Gbps, demonstrating a performance gain of 48 Gbps compared to the baseline. This work provides a relatively comprehensive reference for the design of end-to-end optical interconnection transceivers. Comparison carried out at bitrate above 320 Gbps highlights the scheme's potential for high-capacity optical interconnection scenarios.