O-Band DML Enabled 4×100G Short-Reach Photonic Interconnections With Weighted Decision Aided Recurrent Neural Network

The severe bandwidth constraint along with nonlinearity introduced by the utilization of low-cost direct modulation laser (DML) hinders a higher data throughput for the photonic interconnection. Here, we propose a nonlinear equalizer based on weighted decision aided recurrent neural network (WD-RNN) for the four-level pulse amplitude modulation (PAM-4) intensity modulation direct detection (IM-DD) short-reach transmission, when the average 10-dB bandwidth of four used DMLs is only 20.2 GHz. After optimizing the WD-RNN equalization parameters for one wavelength channel at fixed received optical power (ROP) and applying the same parameters to other channels, we can experimentally realize the 4×100 Gb/s transmission over 10 km standard single-mode fiber (SSMF) and verify the generalization of the WD-RNN enabled nonlinear equalizer under the wavelength channel variation, when the 7% HD-FEC threshold is considered. Furthermore, the proposed nonlinear equalizer is experimentally verified to outperform the conventional model-driven Volterra equalizer, data-driven nonlinear equalizers based on feedforward network, and auto regressive recurrent neural network (AR-RNN), in term of the BER performance under various ROP conditions. Meanwhile, the training process of WD-RNN can be simplified with reduced data size and epochs, in comparison with its AR-RNN enabled counterpart, leading to the mitigation of decision errors arising in the feedback structure. The required computational complexity (CC) of the WD-RNN based nonlinear equalizer is found to be only 11.85%, 26.14%, 71.88% of that for VNLE, FNN and AR-RNN, in terms of achieving the same equalization performance. With the introduction of weight pruning, the CC of WD-RNN can be further reduced by 40%, without any performance penalty.