Static and reconfigurable polarization shaping towards high-power

Vector beams are optical fields with anisotropic polarization profiles that enable tight focusing, multiplexing, and sensing applications, among others. Yet, most generation mechanisms are either alignment-sensitive and lossy, when using interferometry and/ or SLMs, or static when fabricated, when relying on metasurfaces, limiting reconfigurability and high-power use.

In this work, we first benchmark these limitations experimentally: an interferometric SLM setup produced a radially polarized beam but revealed residual ellipticity and strong sensitivity to alignment through Stokes polarimetry, establishing a baseline for improvement. Next, we designed and fabricated an all-silica metasurface to emit radial or azimuthal vector beams. This material is very appealing for high-power applications due to its high damage threshold.


However, analyzing the device revealed pattern merging, and experimental observation showed deviations from the ideal donut, highlighting the challenges of phase accumulation intrinsic to low-index silica. Building on these research steps, we propose a reconfigurable vector-beam generator that pairs a single, disorder-engineered birefringent layer with control of the input's complex amplitude, learned by a polarization-aware, differentiable Fourier-optics simulator. The machine learning algorithm optimizes the input complex amplitude to synthesize arbitrary vectorial targets using the same random medium, maximizing the overlap between them and the output.

By disorder-engineering with grid-search techniques and by backpropagation methods, we obtained overlaps of 0.98, 0.97, and 0.92 for a radial, lemon, and star target, respectively. We were then able to create an algorithm that performs multi-target optimization using a single birefringent layer. This method has the fundamental advantage of being scalable to high-power.