An end-to-end hybrid deep-learning approach for single-shot wavefront sensing and correction

Abstract

Abstract Optical aberrations fundamentally limit the performance of optical systems across scales, from microscopy and astronomy to quantum imaging and ubiquitous technologies such as smartphone cameras, autonomous vehicles, and optical communication networks. Despite sustained progress in wavefront sensing and correction, existing approaches typically require multiple intensity measurements and iterative phase retrieval algorithms and are often restricted to narrowband operation and simple beam profiles. Their performance further degrades when aberrations are weak or obscured by noise. Here we report a single-shot wavefront sensing and correction paradigm that overcomes the limitations of existing approaches. Central to the approach is the integration of a learned optical phase mask that acts as a physical encoder and is jointly optimized with a neural-network decoder. This encoding removes intrinsic phase ambiguities and significantly enhances sensitivity to weak aberrations. By leveraging this hybrid deep-learning architecture, our method directly and unambiguously retrieves Zernike-based phase distortions from a single focal-plane intensity image, eliminating the need for defocus measurements, or iterative phase reconstruction. The framework is intrinsically broadband, exhibits strong robustness to noise, and generalizes across a diverse class of structured light fields. Together, these capabilities establish a scalable and practical foundation for real-time aberration correction in next-generation optical and photonic systems.