Degenerated mode decomposition with convolutional neural network for few-mode fibers. (October 2022)

• Record Type: Journal Article
• Authors: Yan, Baorui; Zhang, Jianyong; Wang, Muguang; Jiang, Youchao; Mi, Shuchao
• Is Part Of: Optics & laser technology. Volume 154(2022)
• Publication Date: 2022-10
• Subjects: Convolutional neural network -- Degenerated mode -- Few-mode fiber -- High-order mode; Optics -- Periodicals; Lasers -- Periodicals; Electronic journals; 621.366
• DOI: 10.1016/j.optlastec.2022.108287
• Languages: English
• ISSNs: 0030-3992

Abstract

Highlights: The deep convolutional neural networks are introduced to perform degenerated mode decomposition. The relationship between the three polarizer-filtered projected intensities and one-dimension modal coefficients are successfully learned. The validation and test results show the feasibility and accuracy of the proposed method. This method has significant potential for waveguide structure field analysis or laser beam characterization. Abstract: In this paper, a deep learning method is proposed to fully characterize the degenerated mode of high-order mode (HOM) group in few-mode fibers (FMFs). The HOM consists of four-fold degenerated spatial modes composed of mode degeneracy and polarization degeneracy. Three polarization projection images are used to recover the modal coefficients. Using the well-trained deep convolutional neural network (CNN) models on randomly generated simulation datasets, a mapping relationship of the two-dimensional intensity distribution to the one-dimensional coefficients space of eigenmodes has been efficiently learned. Two metrics are then used for evaluation on the test samples: (1) the error of the modal coefficients between the original and predicted values; (2) the average image correlation of the original and reconstructed intensity image. The results show that the coefficient errors are only almost one percent while the correlation is up to 99%, which demonstrates the feasibility of the proposed method. In addition, the prediction performance and robustness of the CNN are also assessed based on different image resolutions and different percentages of the neighboring mode group power. The quantitative evaluations demonstrate the stability of the well-trained CNN.