Experimental demonstration of coherent beam combination by a simulation-trained deep neural network

For phase retrieval in a coherent beam combining of 7 fiber amplifiers arranged in a tiled aperture experiment, we demonstrate the feasibility of direct implementation of a light-weight deep-learning model trained on simulated data only. Deep-learning-assisted phase control performs efficiently with lower than λ/30 residual phase error. The use of simulation-trained neural networks allows for fast training (<10 min), a priori optimization, without the need for experimental data acquisition, and easier experimental portability.