Mitigating ambiguity by deep-learning-based modal decomposition method

Abstract

We propose an effective numerical modal decomposition (MD) algorithm for few/multi-mode fibers using a deep convolution neural network (CNN) model in this paper. MD is an available method to reveal modal coefficients. However, with the increase of the superimposed eigenmodes number, the performance of MD will deteriorate due to the modal ambiguity. Our aim is to attain both modal amplitudes and phases from the near-field intensity profile, while minimizing the effect of modal ambiguity as much as possible. Specifically, we train the model with the combination of a modal coefficients loss and two reconstruction losses (near-field and far-field intensity reconstruction losses), which ensures the uniqueness of the solution. With extensive simulational results, we demonstrate that our model is able to mitigate the problem of modal ambiguity and attain accurate modal coefficients (The correlation is above 1.9937 for all modal cases) in a high-speed way. Additionally, the influence of noise and task weights are comprehensively studied. Our proposed technique is useful to mitigate the modal ambiguity.