Deep Learning Methods for Nonlinearity Mitigation in Coherent Fiber-Optic Communication Links

Abstract

Nowadays, the demand for telecommunication services is rapidly growing. To meet this everincreasing connectivity demand telecommunication industry needs to maintain the exponential growth of capacity supply. One of the central efforts in this initiative is directed towards coherent fiber-optic communication systems, the backbone of modern telecommunication infrastructure. Nonlinear distortions, i.e., the ones dependent on the transmitted signal, are widely considered to be one of the major limiting factors of these systems. When mitigating these distortions, we can’t rely on the pre-recorded information about channel properties, which is often missing or incorrect, and, therefore, have to resort to adaptive mitigation techniques, learning the link properties by themselves. Unfortunately, the existing practical approaches are suboptimal: they assume weak nonlinear distortion and propose its compensation via a cascade of separately trained sub-optimal algorithms. Deep learning, a subclass of machine learning very popular nowadays, proposes a way to address these problems. First, deep learning solutions can approximate well an arbitrary nonlinear function without making any prior assumptions about it. Second, deep learning solutions can effectively optimize a cluster of single-purpose algorithms, which leads them to a global performance optimum. In this thesis, two deep-learning solutions for nonlinearity mitigation in high-baudrate coherent fiber-optic communication links are proposed. The first one is the data augmentation technique for improving the training of supervised-learned algorithms for the compensation of nonlinear distortion. Data augmentation encircles a set of approaches for enhancing the size and the quality of training datasets so that they can lead us to better supervised learned models. This thesis shows that specially designed data augmentation techniques can be a very efficient tool for the development of powerful supervised-learned nonlinearity compensation algorithms. In various testcases studied both numerically and experimentally, the suggested augmentation is shown to lead to the reduction of up to 6× in the size of the dataset required to achieve the desired performance and a nearly 2× reduction in the training complexity of a nonlinearity compensation algorithm. The proposed approach is generic and can be applied to enhance a multitude of supervised-learned nonlinearity compensation techniques. The second one is the end-to-end learning procedure enabling optimization of the joint probabilistic and geometric shaping of symbol sequences. In a general end-to-end learning approach, the whole system is implemented as a single trainable NN from bits-in to bits-out. The novelty of the proposed approach is in using cost-effective channel model based on the perturbation theory and the refined symbol probabilities training procedure. The learned constellation shaping demonstrates a considerable mutual information gains in single-channel 64 GBd transmission through both single-span 170 km and multi-span 30x80 km single-mode fiber links. The suggested end-to-end learning procedure is applicable to an arbitrary coherent fiber-optic communication link.