Decoding multimodal spectra of chiral plasmonic nanoantennas using deep learning

Deep learning has emerged as a transformative approach for on-demand inverse design of chiral metamaterials and nanophotonic devices. However, the inherent black-box nature of neural networks (NNs) poses significant challenges in interpreting the multimodal optical responses predicted by data-driven models. We develop a hybrid architecture that integrates an expanding-width NN with a Gaussian parsing NN to achieve spectral decoding and inverse design of chiral plasmonic nanoantennas. The expanding-width NN adopts a neuron-doubling architecture across successive layers, effectively broadening gradient propagation pathways to mitigate vanishing gradients, thereby enabling accurate predictions of nanoantenna scattering, absorption, and near-field spectra (mean squared error <0.0077). Importantly, the Gaussian parsing NN decomposes complex plasmonic spectral responses into fundamental eigenmodes, quantitatively resolving critical parameters such as the dominant mode count, linewidth, and resonance wavelengths. Furthermore, we utilize this framework to establish an end-to-end mapping from geometric parameters to tailored plasmonic modes that govern circular dichroism and Fano resonance. Our methodology not only bridges data-driven design with physical interpretability but also paves the way for new applications in topological photonic circuits, single-molecule biosensors, and hyperspectral computational imaging.