Neural Network Empowered Equalization for SOA Nonlinear Distortions in Coherent Optical Systems

Semiconductor optical amplifiers (SOAs) are promising solutions for small-form-factor optical signal amplification, but their nonlinear gain dynamics and pattern-dependent effects can significantly degrade transmission performance. While digital back-propagation (DBP) techniques have been used to mitigate SOA nonlinearities in coherent systems, they require accurate SOA parameters, which are often difficult to measure. In this work, we examine a SOA booster configuration with a neural network (NN) equalizer that is not based on a physical model and thus does not require knowledge of device parameters. The NN compensates collectively contributions from all sources of nonlinearity, e.g., nonlinear phase noise (NLPN) and chirp. We examine the trade-offs the NN equalization allows between gain, nonlinearity, and launch power by sweeping the SOA injection current and input power. The NN equalization achieves Q-factor improvements of up to 4 dB over linear equalization for 16-quadrature amplitude modulation (QAM) signals at both 30 Gbaud and 60 Gbaud. We demonstrate its effectiveness in compensating SOA distortions over 60 km and 100 km fiber transmission links. The NN equalizer also shows strong robustness to varying training scenarios and signal conditions, and to higher-order modulation formats such as 64-QAM. Performance- complexity analysis of the NN equalizer shows several trade-offs.