Browsing by Author "Uhlemann, Tim"

Now showing 1 - 1 of 1

Open Access
Deep learning for coherent nonlinear optical communications
(2025) Uhlemann, Tim; Ten Brink, Stephan (Prof. Dr.-Ing.)
Communication, in general, is the transport or exchange of information between two geographically distant points. The higher this distance the more sophisticated methods and materials have to be applied to overcome the same. Current state of the art for long-haul communication are optical fibers, that form the crucial backbone of our global, interlinked and digital society, connecting data-centers, factories and homes. Nevertheless, like all other physical media also the optical fiber induces the ever-present attenuation to the information carrying electromagnetic wave what, finally, limits the achievable throughput. This results in the need for higher input powers provided by (laser) diodes. Over the past decades physical and computational limitations led to an operation of the optical fiber in the linear regime, where attenuation as well as dispersion, and, thus, their compensation, constituted the major challenges. As this has changed recently, the investigation of nonlinearity gained more attraction. This work focuses on the pre-distortion and post-equalization of such nonlinear effects that limit the overall efficiency. Thereby, methods from deep-learning are applied and compared to conventional methods. As those, in general, lack of interpretability, here, the concept of so-called architectural templates is proposed that combines well-known, and, from theory derived, concepts with the ones provided by native deep-learning. This way, the results can be analyzed with proven methods from the field of signal processing. While learning of the receiver is a straight-forward, but still complex, task, even more challenging is learning the transmitter as the optimization, i.e., gradient flow, has to be conducted (backwards) through the optical fiber channel. Here, all learnings and evaluations are performed on an accurate simulation of the optical fiber, what enables an isolation of the investigated nonlinear effects. It results that learning over such accurate nonlinear models is possible as the gradient is preserved and the error can be back-propagated. Further, conventional linear filters for dispersion compensation can be outperformed, when being trained in the nonlinear regime. The extension to a nonlinear architectural template revealed the need for a more sophisticated training procedure proposed in this work. When considering a multi-carrier system the trainable nonlinear template for the transmitter was able to exploit additional information from the neighboring channels.