Deep learning based phase retrieval with complex beam shapes for beam shape correction

Beam shaping in laser-based additive manufacturing can enhance printing speed and printed part properties. Complex beam shapes in the Fourier plane are produced by beam shaping systems by modulating the phase in the pupil plane with a phase mask. However, the phase mask may suffer deviations due to heat, etc., leading to a distorted beam shape. This paper provides a solution for phase retrieval with complex beam shapes for beam shape correction. Based on the distorted beam shape, a deep learning model identifies the aberrations represented with Zernike coefficients, which are subsequently used to correct the beam shape. We benchmark the employed computer vision deep learning models against SOTA phase retrieval methods on different beam shapes. Results show that the employed models outperform state-of-the-art methods in aberration detection and beam shape correction on simulated data.