基于神经网络的相位补偿方法

Facing the challenge of slow convergence in traditional adaptive optics (AO) systems, this study proposes a novel phase correction method for laser arrays based on neural networks, utilizing convolutional neural networks (CNNs) for rapid phase prediction and correction. The paper first focuses on a 3×3 laser array, for which a CNN model is trained to process speckle images and predict as well as correct the phase. Experimental results show that this CNN model successfully enhances the mean bucket power (PIB) up to 0.881. For more complex 6×6 laser arrays, dividing them into multiple 3×3 sub-arrays and applying the CNN for compensation allows for effective correction after a single prediction. Compared to traditional Stochastic Parallel Gradient Descent (SPGD) algorithms, the CNN significantly reduces the number of iterations and increases processing speed. Further studies indicate that combining the CNN model with the SPGD algorithm can improve beam quality even more, achieving an iteration speed four times faster than traditional SPGD algorithms. Therefore, this research provides an efficient deep learning-based solution for real-time, high-quality laser array phase compensation, confirming the significant advantages of CNN-based laser array phase correction methods in terms of system real-time performance.