High-Dimensional Diagnostic Encoding with Out-of-Plane Nonlocal Metasurfaces

AI-assisted optical diagnosis requires spectral features that are both information-rich and physically reproducible, yet most nanophotonic biosensors reduce complex optical responses to a single scalar observable, typically resonance shift. Here, we report an out-of-plane nonlocal metasurface for robust, high-dimensional diagnostic encoding. The nonlocal lattice resonance converts refractive-index perturbations into distributed spectral-intensity variations while maintaining strong disorder tolerance, with a wavelength sensitivity to in-plane defects of only 0.1 nm/nm. The out-of-plane geometry further embeds the sensing mode within the substrate, spatially decoupling the optical pathway from the liquid analyte and suppressing absorption- and scattering-induced interference in complex liquids. Combined with deep-neural-network decoding of the full-spectrum response, it improves the diagnostic accuracy from ∼75% for manual scalar classification to 90.0%, with an AUC of 95.0%. These results show that high-dimensional encoding enabled by an out-of-plane nonlocal metasurface can improve AI-assisted diagnostic decoding from complex serum samples.