Deep Learning-Based Bidirectional Design of Localized Surface Plasmons in Nano-Ridge Dimer-on-Mirror Structure

Localized surface plasmons (LSPs) in nanoparticles-on-mirror (NPs-on-mirror) architectures enable extreme field confinement for sensing, optomechanics, and quantum photonics, but inverse design is hampered by complex geometry–spectrum mappings and lack of analytical models. We introduce a data-efficient, physics-guided deep-learning framework for bidirectional design of LSP resonances in gold nanoridge dimer-on-mirror structures. The forward model uses physics-based regularization to improve generalization, achieving test MSE = 2.2×10⁻³ and reducing the generalization gap by >80%. For inverse design we employ a tandem architecture with a dual-objective loss that jointly optimizes spectral fidelity and geometric accuracy to mitigate non-uniqueness in low-data regimes. A grid search of loss weights improves spectral consistency by ≈30% and geometric accuracy by 15.4% versus uniform weighting. Predicted geometries fall within typical fabrication tolerances, offering a physically consistent, fabrication-ready route to scalable nanophotonic device design.