Sensorless Wavefront Reconstruction for Extended Solar Imaging Using a Data–Physics Dual-driven Unsupervised Framework

Abstract

Atmospheric turbulence severely degrades the imaging performance of ground-based solar telescopes, while wavefront sensing for extended solar targets remains challenging because target structures are strongly entangled with turbulence-induced aberrations. To address this problem, this work establishes a robust, target-independent wavefront sensing paradigm that reconstructs optical aberrations directly from extended-scene imagery. We propose a data-and-physics dual-driven unsupervised learning framework that integrates three components: a purposefully constructed adversarial dataset (MOWC-Dataset) designed to facilitate target–wavefront disentanglement through controlled sample pairing, in contrast to conventional random pairing; a Bifurcated-Output Network (BiO-Net) for parallel estimation of Zernike coefficients and target images; and a differentiable optical forward model that enforces physical consistency between predictions and observed far-field images. Experiments demonstrate three major advantages. First, the framework is intrinsically robust to uncalibrated system aberrations, reducing wavefront reconstruction error by more than 70% compared with supervised baselines. Second, it maintains stable performance under noise, with 91.5% of test samples achieving a residual wavefront error below 0.10λ. Third, it exhibits strong “single-model, multitarget” generalization: despite being trained solely on solar imagery without ever seeing nonsolar targets, the model successfully corrects aberrations for structurally distinct nonsolar targets, improving PSNR by 13.20 dB on average and increasing SSIM from 0.290 to 0.906. Overall, this work provides an unsupervised framework for accurate, sensorless wavefront reconstruction of extended targets without phase labels or guide stars, offering a more adaptive paradigm for aberration correction in ground-based solar imaging.