Total: 1
Localized surface plasmons can confine light within a deep-subwavelength volume comparable to the scale of atoms and molecules, enabling ultrasensitive responses to near-field variations. On the other hand, this extreme localization also inevitably amplifies the unwanted noise from the response of local morphological imperfections, leading to complex spectral variations and reduced consistency across the plasmonic nanostructures. Seeking uniform optical responses has therefore long been a sought-after goal in nanoplasmonics. However, conventional probing techniques by dark-field (DF) confocal microscopy, such as image analysis or spectral measurements, can be inaccurate and time-consuming, respectively. Here, we introduce SPARX, a deep-learning-powered paradigm that surpasses conventional imaging and spectroscopic capabilities. In particular, SPARX can batch-predict broadband DF spectra (e.g., 500-1000 nm) of numerous nanoparticles simultaneously from an information-limited RGB image (i.e., below 700 nm). It achieves this extrapolative inference beyond the camera's capture capabilities by learning the underlying physical relationships among multiple orders of optical resonances. The spectral predictions only take milliseconds, achieving a speedup of three to four orders of magnitude compared to traditional spectral acquisition, which may take from hours to days. As a proof-of-principle demonstration for screening identical resonances, the selection accuracy achieved by SPARX is comparable to that of conventional spectroscopy techniques. This breakthrough paves the way for consistent plasmonic applications and next-generation microscopies.