Deep learning-based reconstruction of plasmonic scattering spectra from sparse spectral measurements

Reconstructing scattering spectra from limited spectral measurements is a recurring need for plas-monic nanostructures, both in numerical workflows where full-wave simulations are computationallyexpensive and in experimental settings where broadband scans can be slow, bandwidth-limited, orsparsely sampled. We address this problem with a supervised deep learning framework targeted atgold nanoridge dimer-on-mirror structures. Two complementary masking schemes are studied: acontiguous spectral window, which forces the model to extrapolate beyond the observed region, and arandom sparse mask, which reduces the task to interpolation between scattered observations. Threearchitectures with different inductive biases are compared on a dataset of 1307 COMSOL-simulatedscattering spectra: a multilayer perceptron (MLP), a one-dimensional convolutional neural network(CNN) with dilated kernels, and a one-dimensional U-Net, a U-shaped encoder–decoder networkwith skip connections. We find that no architecture is universally optimal. Under contiguous mask-ing, the MLP outperforms both convolutional models, with R2 ≈ 0.98 from only one third of thespectral range, because its dense connectivity provides a global receptive field suited to long-rangeextrapolation. Under random sparse masking, all three architectures reach R2 ≥ 0.989, but theCNN matches the larger U-Net while using roughly five times fewer parameters. A bandwidth ab-lation in the contiguous case shows that one third of the spectral range is sufficient for high-fidelityreconstruction, with diminishing returns beyond about 40% visibility, and a visibility ablation in therandom case shows that 25–30% of the spectral points is sufficient for reliable reconstruction. Thearchitecture-task coupling identified here gives a concrete design rule for sparse spectral acquisitionand neural post-processing pipelines in nanophotonics.