Time-Domain Learned Digital Back-Propagation

Performance for optical fibre transmissions can be improved by digitally reversing the channel environment. When this is achieved by simulating short segment by separating the chromatic dispersion and Kerr nonlinearity, this is known as digital back-propagation (DBP). Time-domain DBP has the potential to decrease the complexity with respect to frequency domain algorithms. However, when using finer step in the algorithm, the accuracy of the individual smaller steps suffers. By adapting the chromatic dispersion filters of the individual steps to simulated or measured data this problem can be mitigated. Machine learning frameworks have enabled the gradient-descent style adaptation for large algorithms. This allows to adopt many dispersion filters to accurately represent the transmission in reverse. The proposed technique has been used in an experimental demonstration of learned time-domain DBP using a four channel 64-GBd dual-polarization 64-QAM signal transmission over a 10 span recirculating loop totalling 1014 km. The signal processing scheme consists of alternating finite impulse response filters with nonlinear phase shifts, where the filter coefficient were adapted using the experimental measurements. Performance gains to linear compensation in terms of signal-to-noise ratio improvements were comparable to those achieved with conventional frequency-domain DBP. Our experimental investigation shows the potential of digital signal processing techniques with learned parameters in improving the performance of high data rate long-haul optical fibre transmission systems.