D 2 NN-enabled generation of Poincaré sphere vector vortex beams with light-controlled polarization and topological charge

We introduce a novel method for generating Poincaré sphere vector vortex beams (VVBs) via a dual-path diffractive deep neural network (D 2 NN) architecture. This approach enables the simultaneous, integrated control over both polarization state ( α , β ) and topological charge (TC) within a single optical system. Two co-designed D 2 NNs process orthogonal polarization channels, where spatially distinct input regions encode target polarization states (spatial encoding scheme) and source-to-network distances dynamically configure the TC values. Remarkably, a single training cycle enables the generation of 48 distinct VVBs spanning 4 unique Poincaré spheres. Leveraging inherent optical computation, the D 2 NN delivers exceptional fidelity: output purities exceed 99.9%, while polarization angle deviations average only 0.24° ( α ) and 0.8° ( β ). Numerical simulations confirm that performance, including purity and precision, inherently scales with network depth. Addressing fabrication challenges for miniaturized D 2 NNs via 3D printing or femtosecond laser writing, we experimentally validated the concept using a spatial light modulator (SLM) to emulate tailored D 2 NNs. This practical implementation successfully generated 16 VVB types using a simplified two-layer design. This work demonstrates D 2 NNs as a platform for the rapid (light-speed), precise, and versatile transformation of Gaussian beams into diverse targeted VVBs.