Physics-informed deep learning for accurate and efficient wavefront sensing in adaptive optics

Abstract

Accurate and efficient wavefront sensing is central to adaptive optics but remains constrained by the limited receptive fields of convolutional neural networks (CNNs) and the quadratic computational complexity of Transformers. This study introduces state space models to wavefront sensing for the first time, innovatively integrating Zernike-mode point spread functions as physical priors to construct a novel framework. The proposed approach simultaneously achieves global dependency modeling and linear computational complexity. Evaluated on a dataset of 200,000 far-field spot images — spanning 67 Zernike coefficients and diverse atmospheric turbulence conditions — the framework demonstrates significant improvements. Without any pre-training, it reduces test loss and root mean square (RMS) error in aberration estimation by 75% and 52%, respectively, compared to the best CNN baseline. Moreover, it outperforms ImageNet-pretrained Transformer methods while reducing single-frame inference time. Enhanced noise robustness tests further confirm its superior performance and stability. These results validate the potential of physics-informed state space models for high-precision, robust, and rapid wavefront reconstruction. The study also establishes a universal methodological framework for integrating domain knowledge into optical sensing tasks.