Artificial Neural Networks for Self-phase Modulation Compensation in Unrepeated Digital Coherent Optical Systems

Digital coherent systems have revolutionized optical communication networks by dramatically increasing spectral efficiency. However, their maximum capacity is still limited by the combination of noise and nonlinear distortion. To further increase the system capacity, the impact of nonlinear distortion can be mitigated using artificial intelligence. In this work, we apply multilayer perceptrons (MLPs) to reduce the error probability in an unrepeated digital coherent system employing dual polarization and 16-ary quadrature amplitude modulation. We consider two different approaches: on the one hand, an MLP that operates on each polarization independently and, on the other hand, an MLP that processes the two polarizations simultaneously. Numerical results reveal that processing both polarizations leads to better compensation performance since inter-polarization nonlinear crosstalk is partially mitigated. In terms of complexity, however, single polarization processing requires a significantly lower number of operations.