Machine Learning Methods in Coherent Optical Communication Systems

The future demand for digital information will outpace the capabilities of current optical communication systems, which are approaching their limits due to fiber intrinsic nonlinear effects. Machine learning methods promise to find new ways of exploiting the available resources, and to handle future challenges in larger and more complex systems.
The methods presented in this work apply machine learning on optical communication systems, with three main contributions. First, a machine learning framework combines dimension reduction and supervised learning, addressing computational expensive steps of fiber channel models. The trained algorithm allows more efficient execution of the models. Second, supervised learning is combined with a mathematical technique, the nonlinear Fourier transform, realizing more accurate detection of high-order solitons. While the technique aims to overcome the intrinsic fiber limitations using a reciprocal effect between chromatic dispersion and nonlinearities, there is a non-deterministic impairment from fiber loss and amplification noise, leading to distorted solitons. A machine learning algorithm is trained to detect the solitons despite the distortions, and in simulation studies the trained algorithm outperforms the standard receiver. Third, an unsupervised learning algorithm with embedded fiber channel model is trained end-to-end, learning a geometric constellation shape that mitigates nonlinear effects. In simulation and experimental studies, the learned constellations yield improved performance to state-of-the-art geometrically shaped constellations.
The contributions presented in this work show how machine learning can be applied together with optical fiber channel models, and demonstrate, that machine learning is a viable tool for increasing the capabilities of optical communications systems.