We introduce a deep-learning technique to perform complete mode decomposition for few-mode optical fibers for the first time. Our goal is to learn a fast and accurate mapping from near-field beam patterns to the complete mode coefficients, including both modal amplitudes and phases. We train the convolutional neural network with simulated beam patterns and evaluate the network on both the simulated beam data and the real beam data. In simulated beam data testing, the correlation between the reconstructed and the ideal beam patterns can achieve 0.9993 and 0.995 for 3-mode case and 5-mode case, respectively. While in the real 3-mode beam data testing, the average correlation is 0.9912 and the mode decomposition can be potentially performed at 33 Hz frequency on a graphic processing unit, indicating real-time processing ability. The quantitative evaluations demonstrate the superiority of our deep learning-based approach. IntroductionRecently, few-mode fibers (FMFs) have attracted much attention for both fundamental and applied research. Space division multiplexing based on FMFs is a promising way to overcome the anticipated capacity crunch of the single-mode fibers [1]. Larger mode area provided by FMFs helps to suppress the detrimental nonlinear effects and improve the damage threshold, which paves the way to higher power fiber lasers [2]. Furthermore, FMF is a perfect platform for experimental exploration on the complicated spatiotemporal soliton dynamics [3,4] and new nonlinear phenomena [5,6] in multi-mode fibers. With the rapid research progress of FMF, it is highly demanded to characterize the properties of the spatial modes emitting from the FMF, which is named as mode decomposition (MD) technique. With MD techniques, the amplitude and phase information of each eigenmode in the optical fiber can be estimated, providing the complete optical field and the beam properties associated with the field, e.g. wave front [7] and beam propagation factor [8]. Recent years, MD techniques have been widely used in many applications, such as optimizing fiber-to-fiber coupling [9], analyzing mode-resolved gain [10,11] or bend loss [12], diagnosing temporal mode instabilities [13,14], measuring mode transfer matrix [15,16] and realizing adaptive mode control [17,18].In the past few years, various MD methods have been proposed with different techniques, such as spatially and spectrally resolved imaging [19], frequency domain cross-correlated imaging [20], ring-resonators [21], low coherence interferometry [22], correlation filter [23] and digital holography [24]. Although these methods can achieve accurate results, they require consuming post-data processing or intense experimental measurements. Besides these approaches, numerical computing-based MD methods have shown their equal accuracy without complex experimental operations [25][26][27][28]. modes based on weak-guidance approximation [34] and the number of them supported within the fiber depends on the fiber parameters. The purpose of the MD is to predict 2 n ρ and n θ from...

No abstract

High-power mode-programmable orbital angular momentum (OAM) beams have received substantial attention in recent years. They are widely used in optical communication, nonlinear frequency conversion, and laser processing. To overcome the power limitation of a single beam, coherent beam combining (CBC) of laser arrays is used. However, in specific CBC systems used to generate structured light with a complex wavefront, eliminating phase noise and realizing flexible phase modulation proved to be difficult challenges. In this paper, we propose and demonstrate a two-stage phase control method that can generate OAM beams with different topological charges from a CBC system. During the phase control process, the phase errors are preliminarily compensated by a deep-learning (DL) network, and further eliminated by an optimization algorithm. Moreover, by modulating the expected relative phase vector and cost function, all-electronic flexible programmable switching of the OAM mode is realized. Results indicate that the proposed method combines the characteristics of DL for undesired convergent phase avoidance and the advantages of the optimization algorithm for accuracy improvement, thereby ensuring the high mode purity of the generated OAM beams. This work could provide a valuable reference for future implementation of high-power, fast switchable structured light generation and manipulation.

In recent years, machine learning, especially various deep neural networks, as an emerging technique for data analysis and processing, has brought novel insights into the development of fiber lasers, in particular complex, dynamical, or disturbance-sensitive fiber laser systems. This paper highlights recent attractive research that adopted machine learning in the fiber laser field, including design and manipulation for on-demand laser output, prediction and control of nonlinear effects, reconstruction and evaluation of laser properties, as well as robust control for lasers and laser systems. We also comment on the challenges and potential future development.

We incorporate deep learning (DL) into tiled aperture coherent beam combining (CBC) systems for the first time, to the best of our knowledge. By using a well-trained convolutional neural network DL model, which has been constructed at a non-focal-plane to avoid the data collision problem, the relative phase of each beamlet could be accurately estimated, and then the phase error in the CBC system could be compensated directly by a servo phase control system. The feasibility and extensibility of the phase control method have been demonstrated by simulating the coherent combining of different hexagonal arrays. This DL-based phase control method offers a new way of eliminating dynamic phase noise in tiled aperture CBC systems, and it could provide a valuable reference on alleviating the long-standing problem that the phase control bandwidth decreases as the number of array elements increases.