Lightweight neural network for wavefront estimation under long-distance, strong atmospheric turbulence

Wavefront estimation under long-distance, strong atmospheric turbulence remains a critical challenge in free-space optical communication (FSOC). Conventional approaches always suffer from high computational cost and latency. To address this issue, we proposed a lightweight high-precision neural network (LHP-Net), a compact yet accurate model that directly predicts Zernike coefficients from single-frame distorted images under long-distance, strong atmospheric turbulence. The architecture combines an optimized convolutional backbone with a lightweight Zernike-aware attention (LZA) module, enhancing the sensitivity to turbulence-induced aberrations while minimizing computational cost. To rigorously evaluate performance, a large-scale dataset using spectral phase screen simulations was obtained, covering propagation distances up to 10 km and turbulence intensity ranging from weak to strong. Simulation results indicate that LHP-Net achieves up to 92.4% lower prediction error and 37.5% faster inference, exhibiting better performance than a conventional convolutional neural network (CNN). Furthermore, our hybrid training strategy significantly enhances the generalization across different turbulence intensities. Remarkably, LHP-Net maintains robust performance even under extreme turbulence, exhibiting minimal prediction error, providing potential for real-time adaptive optics in next-generation free-space optical systems.