High-accuracy mode decomposition for multi-mode fibers using hybrid network with mini-datasets

A novel mode decomposition method for multimode fiber (MMF) is proposed by using a hybrid network, which combined deep-learning convolutional neural network (DL-CNN) with iterative gradient ascent algorithm (IGAA). DL-CNN is used as the global search for the rough modal amplitudes and relative phases. IGAA is designed as the local optimization to obtain the accurate values of modal decomposition. Although a mini-datasets are employed, the hybrid network shows very good accuracy of modal decomposition and fast convergence. For all of 3-, 5-, 6-, 8-, 10-mode cases, correlation coefficient of the reconstructed and the true near-field intensity patterns can be optimized to higher than 0.98.