Investigation of Convolution Neural Network-based Wavefront Correction for FSO Systems

To explore the effectiveness of the deep learning-based adaptive optics for free space optical communication (FSO), a method of wavefront correction with the AlexNet-based convolution neural network (CNN) is numerically and experimentally investigated. Based on the theory on the Zernike modes (polynomials) for the Kolmogorov turbulence, wavefront aberrations are statistically constructed to mimic the practically received light beams in the weak and strong turbulence. To evaluate this method, comparison in power penalty of a light beam coupled into a standard single mode fiber between experimental and numerical results is discussed. It is shown that the CNN-based wavefront correction significantly improves the power penalty compared with the traditional methods using the stochastic parallel gradient descent (SPGD) and simulated annealing (SA) algorithms in the strong turbulence. From a number of measured light beams with wavefront aberrations, experimental results indicate that average power penalties of 1.8 dB and 0.8 dB are achieved in the strong and weak turbulence, respectively.