Interpretable Deep Learning for Nanophotonics: Feature Importance in Optical Design

Le résumé fourni par la source

Deep neural networks (DNNs) have shown strong potential for modeling the complex optical responses of nanophotonic structures, yet their limited interpretability remains a barrier to adoption in design workflows. This work introduces a dual feature importance (FI) analysis framework-combining weight-based metrics from trained and pruned DNNs with SHapley Additive exPlanations (SHAP) values-to interpret model predictions for multilayer thin films, metasurfaces, and grating nanostructures. Using datasets generated from transfer matrix and finite-difference time-domain (FDTD) simulations, we quantitatively identify the design parameters most influencing optical behavior. Across all structures, FI results align with established physical principles such as diffraction, interference, and resonance. For multilayer films, refractive index and structural periodicity dominate; in metasurfaces, the lattice period and polarization-dependent radii are most critical; and in gratings, groove depth and period are key drivers. SHAP analysis further provides instance-level attribution, complementing global weight-based measures. This interpretable AI framework enhances trust in DNN-driven nanophotonic modeling and offers actionable insights for parameter prioritization in photonic design and optimization.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.