Transfer learning-driven SERS to elucidate mixtures with unknown composition and ratios.

Abstract

Identifying and quantifying compounds in unknown mixtures represents the ultimate goal of surface-enhanced Raman scattering (SERS) spectroscopy but remains a significant challenge in real-world applications. Existing machine learning-driven SERS methods are limited by their reliance on prior knowledge of mixture composition, while time-consuming experimental testing of all possibilities is not feasible. We integrate the molecular specificity of SERS with an adaptive transfer learning (TL) strategy to sequentially identify and quantify carnitine components in 11 unknown binary, ternary, and quaternary multicarnitine mixtures, achieving 100% identification accuracy and a mean quantitation error of only 3%. All models are trained solely on pure compound spectral data, enabling scalable, qualitative, and quantitative analysis of complex, unseen multiplex spectra—without requiring costly and time-consuming training data collection for every possible mixture. This predictive transfer learning-driven approach marks a transformative leap for practical SERS applications, allowing accurate analysis of complex mixtures without prior knowledge of components or ratios.