Hologram generation using attention for Fraunhofer diffraction

Abstract

Controlling optical wavefronts in the strict far-field regime is a fully global inverse problem. Fraunhofer propagation is a single Fourier transform of the aperture, so every modulator pixel couples to every point of the reconstruction and no spatial locality survives for the convolutional priors that dominate learned holography to exploit. Iterative phase retrieval stagnates and trades computation for quality, and camera-in-the-loop methods remain iterative and near-field. Here we show that this global coupling maps naturally onto self-attention, generating a phase-only hologram in a single forward pass of 12 to 24 ms. On hardware, every simulation-trained method collapses to a common floor between 15.8 and 16.1 dB median PSNR set by model-reality mismatch, despite spanning 27.7 to 56.5 dB median PSNR in ideal simulation. Placing the physical optics inside the training loop, which we call self-attention in-the-loop (SAIL), reaches 19.53 dB median PSNR and exceeds the strongest camera-adapted baseline by 0.79 dB on 18 of 18 targets (0.43 dB on 14 of 18 under a more faithful forward model). Most importantly, we show self-attention allows one shared network to serve 18 targets, and as the task complexity increases, the attention mechanism activates. These results establish self-attention as a scalable framework for far-field holography, and attention-based modeling as a general approach to inverse problems governed by long-range interactions in wave physics.