Zero-Shot ML Equalization in Coherent Optical Transmission: A First Experimental Demonstration

There has been a wide range of work demonstrating the effectiveness of machine learning-based equalizers in coherent optical communications. Despite this, the training requirements for applying these techniques remain prohibitive for any real-time implementation. In this paper, we demonstrate a procedure that allows for a neural network equalizer to be trained entirely offline in simulation such that it can be applied, in a zero-shot manner, to a broad range of experimental conditions without the need for online training data. This work therefore presents a technique which may allow for the training requirement of machine learning-based equalizers to be mitigated entirely. In addition, we showcase this technique matching or outperforming conventional linear DSP which needs to be tuned to each channel condition. This also therefore allows for equalization with constant convergence latency when being applied to new channel conditions, enabling fully dynamic networks.