Neural Network based Dynamic Nonlinear MIMO Equalization in 3-D Polarization Multiplexed Direct Detection Systems

This work presents a dynamic nonlinear MIMO equalizer for 3-D polarization multiplexed direct detection systems, which integrates a dynamic MIMO feedforward equalizer (FFE) and a static MIMO neural network (NN). Validated in a 10-km, 150-Gb/s 3-D Stokes vector direct detection (SVDD) experiment, the proposed equalizer yields an increase of 0.1056 normalized generalized mutual information (NGMI) compared with the Volterra equalizer at -20-dBm received optical power. We further compare two NN architectures for multi-dimensional equalization: a single multi-output NN (joint NN) versus multiple single-output NNs (separate NN). It is shown that a separate NN achieves better performance when the signal quality of different dimensions varies. These findings provide useful insights into the design of adaptive equalization for multi-dimensional high-capacity systems.