Launch power profile optimization for band-dependent GSNR in S+C+L WDM transmission systems using autoencoder

Abstract The S+C+L-band transmission system poses a challenge for launch power profile (LPP) optimization due to the significant stimulated Raman scattering (SRS) effects. We propose a digital twin (DT)-based LPP optimization method to precisely achieve band-dependent target generalized signal-to-noise-ratio (GSNR). By generating data through the DT, we train a mirror neural network (NN) and use it as the decoder to construct an autoencoder (AE). A smoothing NN is introduced between the encoder and decoder to ensure the power continuity of adjacent channels. By training the encoder parameters of the AE with the target GSNR, the optimized LPP can be obtained from the output layer of the smoothing NN. We built a simulation setup to optimize the LPPs for two band-dependent target GSNR profiles, achieving average GSNR errors of only 0.24 dB and 0.18 dB across all channels, respectively.