Optical vector-eigenmode decomposition for few-mode fibers through deep learning networks

Abstract

For the optical vector-eigenmode (VE) decomposition, we introduce a novel deep learning-based optical-mode decomposition (OMD) technique for accurate retrieval of near-field beams generated from few-mode fiber (FMF) channel, in which this residual network (Resnet) scheme could obtain complete and accurate amplitude and phase coefficients from the near-field beams. For the case of 3 intra-module modes in the FMF, the correlation coefficient of the OMD reaches 0.9808, while the modal weight error and modal phase error are as low as 0.0287 and 0.0331, respectively. And for the case of 10 inter-module modes in the FMF, the correlation coefficient of the OMD achieves 0.9878, while the mode weight error and mode phase error are recorded down to 0.0071 and 0.0205, respectively. Moreover, the noise-added near-field beams show a good robustness of the OMD based on deep learning. Finally, we also investigate the relationship between modal differences and correlation coefficients, revealing the nature of VE-based OMD functions.