ML-Based Modeling of EDFA Pluggable Modules for OSNR Estimation

Erbium doped fiber amplifiers (EDFA) return an output dependent on the target gain, the channel-loading configuration, as well as the input channels' wavelengths and powers. Due to absence of accurate analytical models, in this article we investigate machine learning methods and their use for estimating the output of an EDFA pluggable module. First, we characterize experimentally a preamplifier. Secondly, after time demanding hyperparameter optimizations we evaluate the performances of deep learning and ensemble methods. Then, we evaluate the generalizability of the models and the benefits of transfer learning strategies to reduce the generalization error. Finally, we propose an algorithm for OSNR estimation that we evaluate based on experimental data.