Seeing Beyond RGB Capabilities: Data-Driven and Physics-Guided Broadband Spectral Extrapolation of Plasmonic Nanostructures by Deep Learning

Localized surface plasmons confine light within deep-subwavelength volumes, enabling ultrasensitive near-field responses that underpin a wide spectrum of interdisciplinary technologies. Yet this extreme localization also amplifies unwanted “noise” from local nanomorphological variations, resulting in spectral complexities and inconsistencies that have long hindered reproducible and scalable nanophotonics. In this context, optical identification and screening of nanostructures with consistent, target responses offers a practical strategy. However, conventional imaging and spectroscopies, including hyperspectral methods, are limited by a resolution–throughput trade-off, motivating the development of faster, high-precision approaches. Here, we introduce SPARX, a deep-learning (DL)-powered paradigm that surpasses conventional imaging and spectroscopic capabilities. SPARX batch-classifies the nanoparticles by their shapes, and extrapolates broadband dark-field spectra (500–1000 nm) of numerous nanoparticles simultaneously from an information-limited RGB image (<700 nm) by learning physical relationships among multiple orders of resonances. Predictions take only milliseconds, achieving a speed-up of 2–4 orders of magnitude over traditional methods, while maintaining comparable precision. This transformative, imaging-DL integrated approach enables reproducible nanoplasmonic applications and, more importantly, fundamentally reshapes optical characterization workflows and extends their reach.