Deep learning-based general beam synthesis for atmospheric propagation

Abstract

Optimizing the transmit light beams unlocks the full potential of free-space optical systems. However, designing application-specific light beams remains a challenge, especially for those traversing random media. In this study, we address this gap by proposing a deep learning-based method to generate optimal beams for propagation through atmospheric turbulence. The key mechanism is approximating the receiver statistics through batch-wise computation during the training of a convolutional neural network (CNN). On that basis, statistical performance metrics including average received power, scintillation index, and mean signal-to-noise ratio (SNR) are considered for optimization. Pseudo-modes of the beam are synthesized by weighted superposition of Hermite-Gaussian eigenmodes, enabling the creation of arbitrary complex amplitude profiles, i.e., general beams. An end-to-end implementation framework is designed to facilitate self-supervised learning and eliminate the need for pre-calculated datasets. Effectiveness of the synthesized beam is validated by wave optics simulation and experiments. In particular, comparison with Gaussian Schell-model beams demonstrates that the synthesized beam can achieve lower scintillation and greater intensity at the same time, leading to markedly enhanced receiver SNR. This advantage persists in a wider range of link configurations, extending the application range of stochastic beams.