Explainable deep learning for inverse design of multifunctional quasi-BIC metasurfaces

Abstract

Quasi-bound states in the continuum (quasi-BIC) dielectric metasurfaces provide exceptional control over subwavelength light–matter interactions. Although deep learning based inverse design alleviates the high computational costs of conventional optimization, its practical reliability is often hindered by the black-box nature of neural networks. Here, we propose an interpretable deep-learning inverse-design framework for a multifunctional all-dielectric silicon nitride eccentric-hole cylinder metasurface. By laterally displacing the inner hole, the unit cell in-plan symmetry is broken, exciting dual quasi-BIC resonances. We also demonstrate a CNN-based inverse-design that retrieves the geometric parameters directly from target reflectance spectra within 63 ms, achieving an R² of 0.998. Furthermore, Integrated Gradients analysis overcomes the black-box limitation by identifying the physical spectral features driving the network's predictions. The designed metasurface demonstrates passive polarization-controlled optical switching and label-free refractive index sensing, achieving a high sensitivity of 432 nm/RIU and a figure of merit of 4888 RIU⁻¹. This explainable AI framework provides a rapid, accurate, and transparent approach for developing advanced high-Q nanophotonic devices.