All-dielectric quasi-BIC metasurfaces integrated with deep neural networks for refractive index sensing

All-dielectric metasurfaces offer enhanced light confinement at the nanoscale with low optical losses. These nanostructures support high-quality-factor (Q-factor) resonances, which are crucial for high-resolution and label-free biosensing. In this work, we propose a highly sensitive refractive index (RI) sensor based on quasi-bound states in the continuum (quasi-BICs) in the metasurface. By introducing an asymmetry in the unit cell, we achieve an ultrahigh Q-factor of 18,430. Numerical simulations demonstrate that the quasi-BIC resonances exhibit strong sensitivity to variations in the surrounding RI, showing an average wavelength sensitivity of 385 nm/RIU over a broad RI range of 1.30–1.50, with a resolution of 2.63 × 10 −4 RIU. The sensor exhibits excellent linearity across this range, which is confirmed by the perturbation theory. To enable real-time and automated RI prediction, we train a deep neural network (DNN) on simulated reflectance spectra of the metasurfaces, achieving a mean squared error (MSE) of 3 × 10 −6 and a prediction accuracy of 0.99. This integrated metasurface–DNN platform can accurately distinguish between normal and cancerous cells and detect urinary tract infections based on RI changes. Due to high sensitivity, linear response, and machine learning integration, the proposed sensor can be used as a compact, label-free, and scalable platform for next-generation biomedical diagnostics.