End-to-end deep learning of geometric shaping for unamplified coherent systems

With the increasing data rate requirements on short-reach links, the recent standardization of unamplified coherent optical systems is paving the way for a cost and power-effective solution, targeting a massive deployment in the near future. However, unamplified systems are introducing new challenges. Particularly, the performance is highly dependent on the peak-to-average power ratio (PAPR) of the transmitted signal, which puts at question the use of the typical constellation formats. In this work, we use an end-to-end deep learning framework to optimize the geometry of different constellation sizes, ranging from 8- to 128-ary constellations. In general, it is shown that the performance of these systems is maximized with constellations whose outer symbols are disposed in a square shape, owing to the minimization of the real-valued PAPR. Following this premise, we experimentally demonstrate that odd-bit constellations can be significantly optimized for unamplified coherent links, achieving power budget gains in the range of 0.5-3 dB through the geometric optimization of 8-, 32- and 128-ary constellations.