Physics-inspired neural network prediction of high-Q near-field spectra for roughness-engineered plasmonic metasurfaces

In plasmonic metasurfaces, structural roughness is conventionally regarded as a source of parasitic loss and degradation of surface lattice resonances (SLRs). Here, we demonstrate that roughness can be strategically harnessed as an active design degree of freedom, orthogonal to conventional geometric parameters. We develop a machine-learning-driven framework that establishes a direct nonlinear mapping from roughness-engineered structural parameters to high-Q near-field spectra. A physics-inspired neural network, integrating a multi-scale attention mechanism with Lorentzian and Fourier decomposition layers, accurately resolves both resonant peak positions and the oscillatory background characteristic of plasmonic responses. Trained on FDTD-generated data, the model predicts near-field spectra and Q-factors of spatially heterogeneous hotspots with coefficients of determination (R 2 ) exceeding 0.90 across all test configurations. Systematic Q-factor mapping across the lattice period-roughness coefficient parameter space reveals a systematic diagonal trend, where the high-Q regime shifts toward larger lattice periods with increasing roughness, reflecting a roughness-modified effective SLR coupling condition. These findings establish engineered roughness as a continuously tunable control channel that grants access to Q-factor regimes unattainable through smooth-structure optimization alone, opening new pathways for the inverse design of metasurfaces operating in subradiant coupling regimes.