Physics-informed PatchGAN for atmospheric turbulence Phase Screen Generation

Cristhof Roosen Runge, Ulisses Dias

DOI: 10.14209/sbrt.2025.1571151291

Evento: XLIII Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2025)

Keywords:

Abstract

We propose a physics-informed generative adversarial network (GAN) for synthesizing Kolmogorov-based atmospheric turbulence phase screens, supporting free-space optical (FSO) link simulation. Our approach integrates a Patch-GAN discriminator with spectral normalization, combined with physics-inspired loss function components, including weighted spectral loss, higher-order moment matching, and minibatch diversity to enhance realism and variability. Training leverages relativistic adversarial loss, instance noise injection, and learning rate balancing for improved stability. Statistical and visual comparisons suggest that our framework produces phase screens with Kolmogorov-like statistics and realistic diversity, offering a tool for modeling optical propagation through turbulence. By providing a differentiable and physically consistent turbulence model, we aim at enabling the integration of realistic atmospheric effects into end-to-end trainable FSO communication systems, thereby facilitating more accurate optimization and overcoming the limitations of non-differentiable classical models.

Download