Diffractive neural networks with polynomial phase masks for laser beam shaping with quasi-continuous diffractive optical elements

We present a novel approach to designing continuous diffractive optical elements (DOEs) for laser beam shaping using diffractive neural networks (DNNs) with trainable polynomial phase masks. This method enables the creation of single phase masks and systems of cascaded phase masks that achieve consistently high beam shaping accuracy, regardless of the initial guess. Additionally, we demonstrate how the approach can be specifically adapted to incorporate the manufacturing conditions of a new type of continuous reflective DOEs, which we verify experimentally in a setup incorporating two DOEs.