Deep learning based wavefront sensing for high quality projection lens from defocused point spread functions

Abstract

One of the options for evaluating the image quality and fine alignment of optical systems is the use of phase retrieval methods. However, real-time alignment is not possible due to high computational complexity and numerical errors. This paper aims to analyze the possibility of real-time quantitative evaluation of the wave aberration function of a designed, high-quality optical system (OS) from a set of noisy PSF images around focus using a convolutional deep learning neural network (DLNN). Working at a wavelength of 405 nm with a numerical aperture of $NA=0.279$, we have prepared a large set of simulated noisy PSF data using the Extended Nijboer-Zernike (ENZ) diffraction theory and performed training and verification of the proposed DLNN models for wave aberration retrieval. The advantage presented on simulated datasets over classical phase retrieval methods is significantly reduced evaluation time due to a pre-trained DLNN and analysis of the accuracy of the diffraction model used. The retrieved aberrations set a limit to the wave aberration function and make the proposed method a suitable candidate for optical workshop metrology, challenging further development.