Multi-dimensional Probabilistic Shaping for Optical Superchannels

We present a multi-dimensional probabilistic shaping algorithm for optical superchannels, complemented with machine-learning assisted soft demapping at the receiver. The proposed system outperforms traditional shaping methods by 0.3-1.05 dB for the same block-length, to enable a 300-km superchannel transmission achieving 800 Gbps data-rate per carrier.