Machine learning-integrated surface-enhanced Raman spectroscopy analysis of multicomponent dye mixtures

Surface-enhanced Raman spectroscopy (SERS) is broadly used in the detection and analysis of materials with its fingerprint-like specificity and high sensitivity. However, resembling signals of analytes highly affect the identification and assignment of spectra, which has become a long-term issue to be solved. In this study, various models of machine learning are utilized and compared to support data analysis of complex SERS spectra. Silver-coated gold core–shell nanocubes (Au@AgNCs) are optimized as SERS substrates for the detection of four common dyes − methylene blue (MB), crystal violet (CV), rhodamine B (RhB) and malachite green (MG). Independent principal component analysis (ICA) was utilized to isolate the signals from the SERS spectra of the dye mixtures, and the isolated signals were further classified by commonly used classification models including K Nearest Neighbors (KNN), Support Vector Machines (SVM), Random Forests (RF), and Convolutional Neural Networks (CNN). The results show that the CNN model achieved an accuracy of 98% in the classification of single dyes and an accuracy of 97% in the classification of dye mixtures, which is significantly better than other models. Based on these findings, we propose ICA combined with CNN-assisted SERS spectroscopy as an effective analytical tool for analyzing dye mixtures.