Diffractive neural networks for robust laser beam shaping with B-spline-parameterized phase masks

Diffractive neural networks (DNNs) represent a powerful framework for laser beam shaping, capable of producing nearly arbitrary three-dimensional intensity distributions by cascading multiple phase masks. However, pixelwise optimization of such phase masks leads to nonideal beam shaping quality and induces high sensitivity to misalignments. Prior work showed that constraining layer training to a polynomial basis, specifically Zernike polynomials, as an intermediate training step leads to improved beam shaping quality while simultaneously reducing misalignment sensitivity, but it hinges on choosing an appropriate polynomial basis and suffers from limited local expressiveness. We instead propose using B-splines, inspired by their application in freeform optics, to represent arbitrary and locally defined continuous surfaces, avoiding the limitations of fixed polynomial bases while retaining their efficiency and robustness. We validate the method experimentally for both 2D and 3D beam shaping with cascaded spatial light modulators (SLMs), while also accounting for practical issues such as pixel crosstalk.