Breaking the Q-limit: wide-range and high-precision metasensing empowered by deep learning

Refractive index sensing traditionally relies on high-Q resonances in precisely fabricated metastructures, making performance vulnerable to fabrication imperfections, limited spectral resolution, and environmental instability. Here, we introduce a fundamentally different paradigm based on computationally learned latent representations rather than engineered photonic sharpness. We experimentally demonstrate an end-to-end variational autoencoder that directly retrieves refractive index from transmission spectra of a generic silicon metasurface, without requiring high-Q features. The model autonomously learns a compact latent manifold encoding refractive-index-dependent spectral information, enabling robust and accurate sensing under strong noise, fabrication variability, and conditions beyond the training distribution. This computational-photonic hybrid approach removes the traditional dependence on resonance finesse and redefines metasurfaces for optical sensing.