Large-dynamic high-accuracy wavefront sensing using deep learning-assisted phase diversity phase retrieval

Phase retrieval (PR), especially under large dynamic aberration conditions, faces challenges such as convergence instability and sensitivity to initial guesses. This paper proposes a novel hybrid approach that integrates deep learning with traditional phase diversity phase retrieval (PDPR) to achieve large-dynamic and high-accuracy wavefront sensing. We introduce a neural network, termed InitNet-PR, which is optimized via neural architecture search based on EfficientNetB0, to provide accurate initial phase estimates from focal and defocused intensity images. These estimates are then used to initialize a pupil-free iterative PDPR algorithm, which avoids reliance on precise pupil amplitude knowledge and enhances applicability in practical optical systems. Simulation results demonstrate that InitNet-PR achieves a residual wavefront RMS of 0.1032λ on the test set, outperforming several benchmark networks. More importantly, when used for initialization, the proposed hybrid method significantly improves convergence probability, reaching 90% even under large aberrations (3.5 ~ 4λ), compared to only 67% with random initialization.