Optimization of fiber optics communication systems via end-to-end learning

One of the key limiting factors in fiber optic communication systems is the nonlinear transmission impairment, due to the Kerr nonlinearity of optical fibers. In order to deal with this impairment, end-to-end learning can be applied to the optical fiber channel. Here, we discuss two different applications.

The first, more standard, approach relies on a standard optical fiber communication system with conventional pulse shaping and makes use of autoencoders to optimize geometric constellation shaping, thus providing a solution robust to nonlinear optical impairments. This method relies on approximate perturbation models of the fiber channel.

In the second approach, instead, end-to-end learning is used to jointly train a nonlinear Fourier transform (NFT) based transmitter and a neural network (NN) based receiver. This second direction therefore makes use of a theoretically optimal signaling scheme for lossless transmission and adapts it to a realistic scenario through end-to-end learning.

In this case, the full split-step Fourier method is used to more precisely model the fiber channel. Through geometric constellation shaping, gains up to 0.13bit/4D numerically and 0.12 bit/4D experimentally can be shown. For end-to-end learning optimized transceivers for nonlinear frequency division multiplexing (NFDM) systems improvement of up to two orders of magnitude are achieved with respect to bit error ratio (BER) in a preliminary numerical analysis.