Deep-learning end-to-end autoencoder for the joint mitigation of chromatic dispersion andKerr nonlinearity in optical communication systems

在这个项目中，我们寻求通过使用人工智能的方法来提高光通信系统的频谱效率，从而增加覆盖范围和数据速率。我们强调了学习的可解释性——通过使用“建筑模板”的概念来促进——从而能够推导出关于光通道及其容量的见解。乍一看，光纤似乎具有无限的带宽，静态传播环境，低噪声和小衰减系数。然而，在过去的几十年里，随着数据速率的指数级增长和电路设计的快速进步，已经把占用的带宽和采样率推向了这样一个运算点：即使是看似完美的光纤也被非线性所支配，这种非线性不再容易被忽视和补偿。从工程的角度来看，这开辟了一个令人兴奋的研究领域，以减轻这种损害。与此同时，深度学习驱动通信已经成为一个非常有前途和活跃的研究课题，特别是在无线领域。已经证明，通过“自动编码器”以联合方式对发射器和接收器进行端到端学习，可以为（几乎）任意信道找到新的信号星座甚至波形，这些信道不限于线性场景，并且以前无法通过经典方法访问。因此，我们被非线性光纤信号传输的挑战和端到端学习框架的概念简单性所吸引。基于我们之前在使用神经网络的无线和光通信方面的成果，我们寻求提出针对光纤通道量身定制的新颖架构模板和学习概念，以提高单波长通道以及基于波分复用的系统的频谱效率、覆盖范围和数据速率。此外，我们打算利用非线性傅立叶变换研究新的结构模板（具有学习到的信号星座和波形）与基于特征值的光通信的关系，以找到联合补偿色散和光纤非线性的进一步思路。

In this project, we seek to improve the spectral efficiency of optical communication systems by using methods from artificial intelligence, and, as a result, increase reach and data rates. We put emphasis on the interpretability of the learnings -- facilitated by using the idea of "architectural templates" -- to enable the derivation of insights about the optical channel and its capacity. At first glance, optical fibers promise a seemingly infinite bandwidth, a static propagation environment combined with low noise and small attenuation coefficients. However, in the previous decades, the exponential growth of data-rates and the fast progress in circuit design have pushed the occupied bandwidth and sampling rates towards an operation point where even the seemingly perfect optical fiber is dominated by nonlinearity that cannot be neglected nor compensated easily anymore. From an engineering perspective, this opens up an exciting field of research to mitigate such impairments. At the same time, deep learning-driven communications has become a promising and active research topic, in particular in the wireless domain. It has been shown that end-to-end learning of transmitter and receiver in a joint manner via an "autoencoder" allows to find new signal constellations and even waveforms for (almost) arbitrary channels that are not restricted to linear scenarios, and that have not been accessible via classic methods before. We, thus, are attracted by the challenges of signaling across the nonlinear optical fiber and the conceptual simplicity of the end-to-end learning framework. Based on our previous results in wireless and optical communications using neural networks, we seek to come up with novel architectural templates and learning concepts tailored to the optical fiber channel, for increasing spectral efficiency, reach and data rates over single wavelength channels, as well as over wavelength division multiplex-based systems. Also, we intend to study the relationship of the novel architectural templates (with learned signal constellations and waveforms) to Eigenvalue-based optical communications using the nonlinear Fourier transformation, to find further ideas for jointly compensating chromatic dispersion and fiber nonlinearities.