Deep Learning-Based Spectral Extraction for Improving the Performance of Surface-Enhanced Raman Spectroscopy Analysis on Multiplexed Identification and Quantitation

Surface-enhanced Raman spectroscopy (SERS) has been recognized as a promising analytical technique for its capability of providing molecular fingerprint information and avoiding interference of water. Nevertheless, direct SERS detection of complicated samples without pretreatment to achieve the high-efficiency identification and quantitation in a multiplexed way is still a challenge. In this study, a novel spectral extraction neural network (SENN) model was proposed for synchronous SERS detection of each component in mixed solutions using a demonstration sample containing diquat dibromide (DDM), methyl viologen dichloride (MVD), and tetramethylthiuram disulfide (TMTD). A SERS spectra dataset including 3600 spectra of DDM, MVD, TMTD, and their mixtures was first constructed to train the SENN model. After the training step, the cosine similarity of the SENN model can achieve 0.999, 0.997, and 0.994 for DDM, MVD, and TMTD, respectively, which means that the spectra extracted from the mixture are highly consistent with those collected by the SERS experiment of the corresponding pure samples. Furthermore, a convolutional neural network model for quantitative analysis is combined with the SENN, which can simultaneously and rapidly realize the qualitative and quantitative SERS analysis of mixture solutions with lower than 8.8% relative standard deviation. The result demonstrates that the proposed strategy has great potential in improving SERS analysis in environmental monitoring, food safety, and so on.