Investigation of multivariate analysis of surface-enhanced Raman scattering spectra using simple machine-learning models: prediction of the composition of mixed self-assembled monolayer on gold surface

Abstract This study investigates the multivariate analysis of surface-enhanced Raman scattering (SERS) spectra using simple machine-learning models. Substrates with gold nanostructures whose surfaces contain self-assembled monolayers (SAMs) of mixed benzene thiol derivatives (as model molecules) are fabricated, and their SERS spectra are acquired. After preprocessing, the spectra are analyzed using different machine-learning models to compare the prediction accuracies of the mixing compositions of the SAM reagents. The results show that linear discriminant analysis (LDA) achieves the highest prediction accuracy for each combination. The accuracy of LDA (0.996) is higher than that of linear analysis with one variable (0.656), which is conventionally employed in analytical chemistry. The prediction accuracy is 0.994 even when the spectral data of three different SAM reagents are combined into a single data-set, which is used for predictions. This suggests that the model exhibits high performance even when the number of prediction categories increases. We analyze the SERS spectrum with a simple model such as LDA and achieve highly accurate predictions of the mixing compositions of the SAM reagents. This indicates that, even when the measurement data of the SERS spectra are scarce, highly accurate predictions can be made through analysis using a simple prediction model.