With increasingly higher data rate requirements imposed on short-reach links, coherent optical systems are expanding to shorter distances. To achieve a cost-efficient and low-complexity solution, optical amplifiers can be removed, enabling the deployment of the already standardized use of unamplified coherent links. However, the system performance is now governed by a peak power constraint, and therefore the conventional constellations become sub-optimal.
We will present how an end-to-end machine learning algorithm can optimize the constellation geometry. We also show the importance of a well-suited model for the channel, by experimentally comparing the performance achieved with geometries optimized for both amplified and unamplified links.