Modeling the second stage of extended L-band fiber amplifiers using neural networks trained on experimental data

Neural networks are fast and accurate in modeling L-band erbium-doped fiber amplifiers with a single stage. To extend this approach to second-stage (or mid-stage) amplifiers, we must address nonuniform and high-power input signals, as well as the presence of significant amplified spontaneous emission in the L-band. We present an experimental method to collect a large training set (15,000 points) for a neural network (NN) that can capture the behavior of gain and noise figure in second-stage amplifiers. We demonstrate that our neural network model has average error below 0.27 dB for gain, or 0.15 dB for noise figure. We examine strategies for collection of training sets, especially in terms of the granularity of the fiber lengths. To show the utility of the NN model as a design tool, we use it to optimize the mid-stage filter of a fixed double-stage amplifier through particle swarm optimization. We contrast mid-stage filters that target 1) flat gain and 2) low noise figure.