AI-Maxwell inverse design for broadband EUV attosecond metasurfaces

Vacuum-guided extreme-ultraviolet (EUV) metasurfaces offer a route to wavefront control in a spectral region where conventional transmissive optics are strongly constrained. Extending this concept from narrowband focusing to broadband attosecond operation requires simultaneous control of spectral phase, focal-energy concentration, temporal fidelity, and sensitivity to nanometre-scale geometric perturbations. Here we identify a robust phase-coverage–energy-stability conflict as a governing bottleneck in a silicon aperture-membrane platform. The design workflow combines periodic RCWA library construction, perturbation-aware admission, complete-aperture reduced-order propagation, pulse reconstruction, and conditional geometry generation. After the final aperture assignment was established, heterogeneous 7 × 7 finite-cluster FDTD calculations were performed for five high-usage selected cells as post-design local consistency checks. These calculations examine local field behaviour and phase-order preservation in model heterogeneous neighbourhoods, but they do not participate in metaatom admission or device optimization. The selected design reaches a mean first-Airy-radius encircled-energy fraction of 0.546, a minimum encircled fraction of 0.473, a maximum FWHM-to-diffraction-reference ratio of 0.987, and a pulse-to-transform-limit ratio of 0.981. Conditional generation supplies independently RCWA-screened candidates that enter the device assignment but do not raise the saturated energy metric. The results identify a platform-specific trade-off between robust phase-state availability and transmitted-energy stability and establish an “AI proposes, Maxwell decides” workflow with clearly separated design, admission, device evaluation, and post-design local validation stages.