Integrating machine learning with surface-enhanced Raman scattering (SERS) accelerates the development of practical sensing devices. Such integration, in combination with direct detection or indirect analyte capturing strategies, is key to achieving high predictive accuracies even in complex matrices. However, in-depth understanding of spectral variations arising from specific chemical interactions is essential to prevent model overfit. Herein, we design a machine-learning-driven "SERS taster" to simultaneously harness useful vibrational information from multiple receptors for enhanced multiplex profiling of five wine flavor molecules at parts-per-million levels. Our receptors employ numerous noncovalent interactions to capture chemical functionalities within flavor molecules. By strategically combining all receptor-flavor SERS spectra, we construct comprehensive "SERS superprofiles" for predictive analytics using chemometrics. We elucidate crucial molecular-level interactions in flavor identification and further demonstrate the differentiation of primary, secondary, and tertiary alcohol functionalities. Our SERS taster also achieves perfect accuracies in multiplex flavor quantification in an artificial wine matrix.
Keywords: chemometrics; flavor analysis; machine learning; molecular receptor; multiplex detection; support vector machine (SVM); surface-enhanced Raman scattering (SERS).