Deep learning-driven spectral exploration for metal-insulator-metal plasmonic metasurfaces

Metal-insulator-metal plasmonic metasurfaces exhibit intricate spectral responses arising from the interplay among localized surface plasmon polaritons, surface lattice resonances, and Fabry-Pérot cavity modes. However, traditional characterization methods relying on iterative electromagnetic simulations and manual spectral analysis face inefficiencies in handling complex parameter spaces and measurement-condition heterogeneity. Here, we present a deep learning-driven framework to analyze the spectral behaviors of metal-insulator-metal metasurfaces by integrating experimental fabrication, finite-difference time-domain simulations, and data-driven spectral classification and regression. Gradient-parameter metasurfaces with varying insulator gaps (20-200 nm), nanostructure geometries (disc/ring), and periodicities (500-1500 nm) are fabricated via electron-beam lithography and optically characterized under reflection/transmission configurations. Numerical simulations reveal the interplay of hybridized modes and surface charge dynamics. Leveraging convolutional and recurrent neural networks (CNNs, LSTMs, GRUs) and Transformers, we achieve robust classification of 24 structural categories and spectral regression for inverse design. Notably, LSTM models attain superior classification accuracy (>99.2%), while CNN demonstrates superior time efficiency. This work establishes a data-physics-integrated paradigm for rapid MIM metasurface characterization, and the proposed methodology bridges the gap between complex optical responses and deep learning-driven spectral exploration, advancing applications in label-free biosensing and tunable photonic systems.