Abstract
Modern optical communication systems exhibit rapid growth due to the advent of edge-cloud networking architectures that fuel the future of 5G/B5G define how internet will evolve in the next decades. As transmission rates continue to grow approaching the physical limits of single-mode fiber, the goal of high capacity without substantially increasing power consumption is a significant challenge. In long haul systems, where digital coherent solutions prevail, the improvement of spectral efficiency is mainly hindered by the nonlinear effects attributed to the Kerr effect. In intensity modulation/direct detection (IM/DD) systems the main degradation comes from power fading, caused by chromatic dispersion, bandwidth limitations and transceiver nonlinearities. The evolution of machine learning algorithms and deep learning techniques makes it possible to apply them to state-of-the-art digital processors in order to efficiently solve complex nonlinear problems. Hence, machine learning becomes a ...
Modern optical communication systems exhibit rapid growth due to the advent of edge-cloud networking architectures that fuel the future of 5G/B5G define how internet will evolve in the next decades. As transmission rates continue to grow approaching the physical limits of single-mode fiber, the goal of high capacity without substantially increasing power consumption is a significant challenge. In long haul systems, where digital coherent solutions prevail, the improvement of spectral efficiency is mainly hindered by the nonlinear effects attributed to the Kerr effect. In intensity modulation/direct detection (IM/DD) systems the main degradation comes from power fading, caused by chromatic dispersion, bandwidth limitations and transceiver nonlinearities. The evolution of machine learning algorithms and deep learning techniques makes it possible to apply them to state-of-the-art digital processors in order to efficiently solve complex nonlinear problems. Hence, machine learning becomes a significant candidate for deal with nonlinear transmission impairments in optical fibers. The main results of this thesis were the treatment of optical fiber nonlinearities through the application of recurrent neural networks (RNN) and the emulation of photonic systems using machine learning algorithms. The use of bidirectional RNNs brings about the significant improvement of the performance of high-speed optical communication systems. The thesis was based on analyses, simulations and experiments carried out in various application scenarios, both in long haul networks (>1000 km) using a coherent receiver and in short reach optical networks (< 120 km) based on IM/DD systems. The most important topics analyzed during the preparation of this thesis are the following: •The use of a Long Short Term Memory Neural Network (LSTM) to compensate for optical fiber nonlinearity in digital coherent systems was presented for the first time. Numerical simulations of C-band and O-band transmission systems for single channel and multi-channel 16-QAM modulation format with polarization multiplexing were performed. The effect of the number of hidden units and the length of the word of symbols were analyzed in detail. The performance and operating limits were extensively studied while the robustness of the proposed equalizer in various scenarios of channel variability was highlighted. The simulations proved that LSTM-type neural networks can be very efficient as digital processors, providing superior performance and at the same time are less complex compared to Digital Back Propagation (DBP) equalizers, especially in the long-distance multi-channel scenario ( > 1000 km). •The complexity and performance of processing units for the removal of non-linear effects were studied. The dominant types of RNN, bi-LSTM, bi-GRU and bi-Vanilla-RNN were compared in terms of complexity and performance, and it was shown that all have the potential to achieve nonlinear compensation especially in digital coherent systems with no in-line dispersion compensation, which carry 16 and 32 QAM modulation signals, with polarization multiplexing, single or multi-channel. Simulations show that the three models provide similar compensation performance, so in real systems, the simpler scheme based on the Vanilla-RNN model should be preferred. The many-to-many approach was proposed as a particularly efficient method of minimizing complexity. A comparison with non-linear Volterra-type equalizers was performed and the superiority of Vanilla-RNN was demonstrated in terms of both performance and complexity. •Based on an experimental 4 × 50-Gb/s IM/DD O-band CWDM system, the significantly better BER performance was demonstrated of the Vanilla-RNN scheme versus the conventional decision feedback equalizer (DFE) for OOK and PAM4 configurations. The ability of the Vanilla-RNN equalizer to compensate for both linear and non-linear impairments induced by the transmitter and single-mode fiber was demonstrated. As a result, optical transmission of up to 100 km and 75 km for the OOK and PΑM4 modulation formats respectively can be realized. Furthermore, through comparison with other equalization schemes including the linear equalizer, the 3rd-order Volterra equalizer and the Volterra+DFE, it was shown that the Vanilla-RNN equalizer achieves the best performance. •Multi-channel equalization of WDM channels was proposed and implemented in order to compensate for the non-linear effects between channels by suitably modifying the bi-RNN processors of the previous works. The equalizer used the multi-input multi-output (MIMO) processing architecture to decode the WDM channels. In this way, it exploited useful information from neighboring channels to better identify and remove inter-channel crosstalk. It was demonstrated through extensive numerical simulations and experimental results that the proposed approach outperformed both standard multi-channel equalization in the form of adaptive equalizers, DBP-type equalizers, and single-channel bi-RNN. •The previous RNN structures were used to emulate the dynamic behavior of a real-life 106.25 Gb/s PAM4 optical transmission system based on a vertical cavity surface-emitting laser (VCSEL). It was shown that RNN networks are capable of reproducing the dynamic characteristics of the optical system with a prediction accuracy approaching 100%. Finally, the ability to emulate functional areas of the system in which the neural network has not been trained was studied.
show more
