Phase Retrieval for Optical System Characterization: Design, Performance Analysis, and Experimental Validation of a Convolutional Neural Network

We introduce a Convolutional Neural Network (CNN) designed for precise wavefront retrieval from point-spread function (PSF) intensity images. Our ResNet18-based CNN infers the first 15 Zernike standard coefficients using three PSF measurements symmetrically positioned around the focal point. The CNN is trained on a dataset of 300,000 simulated PSF image sets containing 4th and 6th order aberrations, with wavefront amplitudes of up to 10 λ. We achieve an accuracy exceeding 99% for each individual Zernike coefficient, with uncertainties ≤ λ/30, as validated on a dataset of simulated PSF images. We evaluate the CNN's performance on experimental PSF measurements, and the predictions are compared to direct wavefront measurements from a Shack-Hartmann sensor. The results indicate prediction accuracy better than λ/15 for each of the 15 coefficients. This confirms the CNN's potential for characterizing optical systems with complex aberration distributions of low to moderate amplitudes.