Optical Field-to-Field Translation under Atmospheric Turbulence: A Conditional GAN Framework with Embedded Turbulence Parameters

We propose a field mapping approach for the propagation of laser beams through atmospheric turbulence, leveraging a Generative Adversarial Network (GAN). The proposed GAN utilizes a U-Net architecture as its generator, with turbulence characteristic parameters introduced into the bottleneck layer of the U-Net, enabling effective control over the generator. This design allows for the flexible simulation of Gaussian beam propagation across a range of turbulence intensities and transmission distances. A comparative analysis between the neural network predictions and numerical simulation results indicates that the neural network can achieve a field mapping speedup of four orders of magnitude while maintaining a relative error within 16% for the second-order statistical moments of the light spot. Additionally, the study investigates the effect of varying turbulence intensities on prediction accuracy. The results indicate that high-frequency speckle patterns caused by beam breakup are the primary factor limiting prediction accuracy under strong or saturated turbulence conditions.