Chaotic optical communications were originally proposed to provide high-level physical layer security for optical communications. Limited by the difficulty of chaos synchronization, there has been little experimental demonstration of high-speed chaotic optical communications, and point to multipoint chaotic optical networking is hard to implement. Here, we propose a method to overcome the current limitations. By using a deep-learning-based scheme to learn the complex nonlinear model of the chaotic transmitter, wideband chaos synchronization can be realized in the digital domain. Therefore, the chaotic receiver can be significantly simplified while still guaranteeing security. A successful transmission of 32 Gb/s messages hidden in a wideband chaotic optical carrier was experimentally demonstrated over a 20 km fiber link. We believe the proposed deep-learning-based chaos synchronization method will enable a new direction for further development of high-speed chaotic optical communication systems and networks.

In this paper, transmission performances of directly modulated laser (DML), electro-absorption modulated laser (EML) and Mach-Zehnder modulator (MZM) are experimentally compared in dispersion-unmanaged high-speed transmission systems with digital signal processing (DSP). We show that, although the DML based transmitter is often believed to be less favorable in C-band high-speed transmissions, it exhibits superior performance over the other two transmitters when either linear or nonlinear digital signal processing is adopted. By theoretical and experimental analysis, we reveal that the superiority of DML can be attributed to the compensation of fiber power fading by its inherent adiabatic chirp as well as the mitigation of chirp induced distortions by the linear or nonlinear equalization. Experimental results of 56Gb/s 4-level pulse amplitude modulation (PAM4) signals under various equalization schemes including linear feedforward equalization, simplified nonlinear Volterra equalization and partial response signaling are presented. Particularly, we show that for DML a 40km transmission distance can be achieved to satisfy the extended range-4 (ER4) Ethernet interconnect using a simplified Volterra equalizer, and a 20km transmission distance can be supported using a linear equalizer. In contrast, for MZM and EML, the achievable transmission distances are respectively 20km and 15km using the Volterra equalizer, respectively, and 15km and 10km using linear equalizer, respectively. Moreover, we show that even using the combination of the Volterra equalizer and partial response signaling, the transmission distances of MZM and EML based systems are limited to 30km and 20km.

Fiber nonlinear interference (NLI) modeling and monitoring are the key building blocks to support elastic optical networks. In the past, they were normally developed and investigated separately. Moreover, the accuracy of the previously proposed methods still needs to be improved for heterogenous dynamic optical networks. In this paper, we present the application of machine learning (ML) in NLI modeling and monitoring. In particular, we first propose to use ML approaches to calibrate the errors of current fiber nonlinearity models. The Gaussian-noise model is used as an illustrative example, and significant improvement is demonstrated with the aid of an artificial neural network. Further, we propose to use ML to combine the modeling and monitoring schemes for a better estimation of NLI variance. Extensive simulations with 2411 links are conducted to evaluate and analyze the performance of various schemes, and the superior performance of the ML-aided combination of modeling and monitoring is demonstrated.

Reservoir computing (RC) by supervised training, a bio-inspired paradigm, is gaining popularity for processing time-dependent data. Compared to conventional recurrent neural networks, RC is facilely implemented by available hardware and overcomes some obstacles in training period, such as slow convergence and local optimum. In this paper, we propose and characterize a novel reservoir computing system based on a semiconductor laser with double optoelectronic feedback loops. This system shows obvious improvement on prediction, speech recognition and nonlinear channel equalization compared to the traditional reservoir computing systems with single feedback loop. Then some influencing factors to optimize the performance of the new RC are numerically studied, and its great potential of addressing more complex and troubling problems in information processing is expected to be exploited.

Chaotic optical communication was originally proposed to provide high-level physical layer encryption. For high-speed and long-distance transmission, chaotic signal is very sensitive to channel impairments such as dispersion and Kerr fiber nonlinearity. In the traditional chaotic optical communications, these impairments must be compensated in optical domain before chaos synchronization. However completely compensating the high-order dispersion and fiber nonlinearity in the optical domain has great challenges, therefore limiting the transmission distance of high-speed chaotic optical communications less than 150 km. Here we propose a method aiming to break the limit. Thanks to coherent detection, channel impairments can be compensated in the digital domain using various algorithms. Digital backpropagation algorithm is used for jointly compensating linear and nonlinear impairments of the chaotic signals, constant modulus algorithm is used for channel equalization and extended Kalman filter is adopted for carrier phase recovery. After digital processing, the recovered chaos signal is converted back to the optical domain for chaos synchronization. By this means, we demonstrate a 10 Gb/s phase-modulation signal encrypted by phase chaos transmission over record-breaking 1000 km single-mode fiber with bit-rate error less than 1×10 $^{-3}$ by simulation, and support the results with extensive numerical analysis.

To support the development of intelligent optical networks, accurate modeling of the physical layer is crucial. Digital twin (DT) modeling, which relies on continuous learning with real-time data, provides a new paradigm to build a virtual replica of the physical layer with a significant improvement in accuracy and reliability. In addition, DT models will be able to forecast future change by analyzing historical data. In this tutorial, we introduce and discuss three key technologies, including modeling, telemetry, and self-learning, to build a DT for optical networks. The principles and progress of these technologies on major impairments that affect the quality of transmission are presented, and a discussion on the remaining challenges and future research directions is provided.