Physics-Consistent Generative Hologram Design for Multifocal Femtosecond Laser Fabrication

Abstract Multifocal phase holography offers a route to parallel femtosecond laser micro- and nanofabrication, but conventional weighted Gerchberg–Saxton (GSW) phase retrieval requires target-specific iterations and does not jointly optimize the optical quantities governing exposure quality. Here we introduce PhyFlow, a physics-consistent invertible flow model that generates phase-only multifocal holograms from target multifocal patterns in a single inverse pass. A multiscale invertible backbone learns target-to-phase correspondence, while differentiable Fourier optics constraints optimize diffraction efficiency, focal uniformity, and target-shape preservation. Across focal numbers and array scales, PhyFlow matches or surpasses GSW in energy concentration, focal balance, and shape fidelity while reducing per-hologram generation time by 98.8%. Optical-field measurements and femtosecond laser writing further show improved dot-area consistency, narrower line width distributions, and fewer nontarget written features. PhyFlow therefore connects rapid learned hologram generation with physically optimized fields and more uniform parallel fabrication.