Deep learning-driven full complex model of highly multimode fibers

We present an original deep learning process for modeling a multimodal fiber by use of a neural network. It connects the complex incident field on a fiber with a large number of modes (140 LP modes per polarization) to the output one, without any interferometric setup or reference beam. Using this model, we demonstrate a remote 2D shaping of the field from the multimode fiber, proving the accuracy of our full complex neural network model.