Enhancing Machine Learning Based Physical Layer Performance Estimation of Optical Transmission Links Using Data Augmentation

Machine Learning (ML) methods can enhance the accuracy of the physical layer performance estimations for optical links, especially in cases where the analytical expressions fall short. This improvement can lead to relaxed operational margins, enabling greater transmission distances and/or increased network capacity. The benefits are most pronounced when ML models are trained on large datasets and for this purpose, in this work, three techniques are introduced that can increase the size of the training set by several orders of magnitude. The results demonstrate that by using these techniques, the average estimation error can be decreased by 1.5 dB, showing significant gains, especially when the initial dataset is small. Moreover, the application of these approaches can also reduce the number of cases where the discrepancy exceeds 2 dB.