
WP1 Development of new NFT transmission methods (Aston University)
The existing NFT spectrum modulation techniques have been reviewed by ESR1 (V. Neskorniuk), who inferred that the following two methods have the potential to render the highest efficiency: b-modulation allowing us to attain the control over the signal’s duration, and the periodic NFT, which can bring about benefits in signal processing and signal-noise interference.
Within WP1 ESR1 developed a new data-driven approach (neural networks-based, NN) to nonlinearity mitigation in optical fibre links, addressing, specifically, the regime of high nonlinearity. He showed that the NN is able not only to recover the nonlinear impairments caused by optical fiber propagation but also the imperfections resulting from the usage of low-cost legacy transceiver components, such as digital-to-analog converter and Mach-Zehnder modulator.
WP2 Impact of practical impairments on the NFT (TU Delft)
Nonlinear distortions limit the transmission capacities of current fiber-optic communication systems. Non-conventional transmission techniques based on nonlinear Fourier transforms (NFTs) are an interesting approach to address these issues. These transmission techniques, however, rely on an ideal lossless fiber model, while in practice fibers are lossy. ESR2 (V. Bajaj) explored the impact of the fiber-loss on NFT-based fiber-optic transmission systems. He also investigated a novel approach to incorporating loss into NFT-based systems that uses a modified NFT and operates in specialized dispersion decreasing fibers (DDFs). The combined use of DDF along with the modified NFT eliminates the degrading effects due to fiber-loss completely in a mathematical ideal scenario. The numerical assessments showed significant gains achieved using the proposed solution.
WP3 Machine learning techniques for fibre-optic channel (DTU)
At Technical University of Denmark (DTU), the ESR3 (S. M. Ranzini), together with the industry partner (NBL), is proposing and developing a new receiver based on optoelectronic machine learning for intensity-modulated and direct detection systems. The optical pre-processing stage slices the received signal spectrum in small sub-bands with passive optical filters and each is detected by a photodetector (PD). The digital post-processing is based on a recent technique in machine learning called reservoir computing. ESR3 demonstrated the potential of the receiver for 32-GBd OOK signal transmission, and showed an increase in reach from 10 km to 40 km, corresponding to 400%, compared with digital-only techniques. The details of this process were published in 2 journal publications and 1 conference paper.
WP4 Network applications of NFT technology (Telecom Paris)
End-to-end deep learning of the optical fiber channel has recently been proposed to address the limitation that the Kerr nonlinearity sets on the transmission rates of fiber optic communication systems. It is important to understand how this approach compares with the conventional methods. By designing a neural network approximating the channel, ESR4 (A. Shahkarami), together with his supervisors, studied this comparison for a small-scale system, which we are currently extending to large-scale systems. In addition, they carried out some research on an approach based on representation learning and feature transfer to help protect the sequence of symbols at the transmitter against errors introduced by the channel.
WP5 Experimental implementation and testing of NFT systems (Nokia Bell Labs)
The performance of coherent optical high-speed transceivers are limited by their physical limitation and impairments. To operate with their maximum capacity, we should mitigate undesired distortions of these devices. A cost-effective way to overcome this challenge is using digital pre-distortion (DPD) techniques. ESR2 investigated a neural network DPD technique and showed in lab experiments an improvement of 3 dB compared to traditional methods. The ESR3 is investigating a new transceiver based on sharing the complexity between the optical and digital domain with machine learning techniques. Experimental analyses were carried out at NBL and showed a transmission reach gain of 800%, compared to digital-only techniques.
Progress beyond the state of the art and expected potential impact
Addressing the development of new data-driven approaches to nonlinearity mitigation, the new neural-networks based equaliser developed within WP1 is beyond state-of-the-art, according to the quality of the signal recovery and its ability to deal simultaneously with the fibre-nonlinearity induced. The important property of the method developed is that it is adaptive, so can be potentially implemented in real metro links resulting in the improved quality of transmission. This is especially important in view of the mitigation of the COVID-19-related negative socio-economic impact and further long-lasting consequences, when we have converted into on-line work in most aspect of our live, such that most business and public activities have become essentially dependant on the speed and quality of virtual information exchange.