Shallow Neural Network Boosts the Evaluation of OAM Fibers

Over the past few years, various optical fibers have been proposed for the generation, transmission, and amplification of orbital angular momentum (OAM) beams. To evaluate the optical properties of these OAM fibers under different fiber parameters, traditional methods usually require much time and effort to solve the Maxwell's equations. In this paper, for the first time, we introduce a single-hidden-layer neural network (NN) to efficiently evaluate OAM fibers. This shallow NN can learn the mapping from the input fiber parameters to the output OAM properties with 0.1% samples generated by traditional methods. Then the NN can fast and accurately evaluate the OAM fibers for the rest samples without the need to solve the Maxwell's equations. The proposed approach takes only about 0.07 ms to evaluate the OAM properties, which is four orders of magnitude faster than traditional methods. Besides, the average evaluating error is smaller than 0.11%. More interestingly, we find the NN can identify and correct the wrong evaluation from traditional methods. The results show that the shallow NN paves the way to a superfast, accurate, and robust evaluation of OAM fibers.