Wavefront prediction using frequency graph neural network for adaptive optics

Adaptive Optics (AO) technology is an effective method for compensating atmospheric turbulence. However, inherent system delays restrict the correction performance, preventing real-time compensation. To overcome these limitations, forward prediction of atmospheric turbulence has become crucial. In this study, we propose a spatiotemporal prediction model based on the Frequency Graph Neural Network (FGNet) to address the challenge of accurately predicting random atmospheric turbulence. FGNet embeds Zernike coefficients, representing multiple frames of wavefronts, into a hypergraph, with each coefficient treated as a graph node. Meanwhile, we transform the temporal domain into the frequency domain, utilizing graph Fourier operators to capture the spatiotemporal features of the data. Compared to traditional AO methods and LSTM-based deep learning models, our proposed FGNet achieves superior prediction accuracy and stability on simulated data, offering an innovative and efficient solution for real-time atmospheric turbulence correction.