ABSTRACT Atmospheric turbulence‐induced wavefront aberrations significantly degrade performance in optical imaging and laser transmission systems. While adaptive optics (AO) offers compensation, conventional systems rely on wavefront sensors and guide stars, limiting their applicability in complex scenarios. Deep learning‐based approaches have emerged as promising alternatives but remain constrained by ill‐posedness and limited nonlinear representation capabilities. To overcome these challenges, we propose a novel dual‐plane nonlinear wavefront sensing method WaveKAN (WKAN) that operates without wavefront sensors or guide stars during inference and deployment. By capturing on‐focus and defocused images to constrain the solution space effectively, WKAN incorporates learnable activation functions and multi‐head self‐attention to enhance its ability to approximate the complex nonlinear mapping and model cross‐scale aberration features, respectively. Validated under various turbulence conditions using both spots under laser and extended objects under LED, WKAN demonstrates superior wavefront reconstruction accuracy (∼0.3 rad on average) and generalization capability compared to existing methods. In wavefront correction experiments, it significantly increased the Strehl ratio for point sources and restored distorted images to be clear for extended objects. These results confirm the potential of WKAN as a robust, guideless solution for broadening AO applications.
Feng et al. (Sat,) studied this question.