Quantitation of surface-enhanced Raman spectroscopy based on deep learning networks

Abstract

Surface-enhanced Raman scattering (SERS) is a highly sensitive detection method that is widely applied in numerous fields. However, the distribution of SERS "hotspots" and their sensitive response at the nanoscale render the reproducibility and quantitative analysis of SERS spectra difficult. In this study, an analytical method based on deep learning was applied for the quantitative detection of SERS spectra. Using Ag/TiO2 composite nanofilms as SERS substrates, the SERS spectra of Rhodamine 6G (R6G) at concentrations of 10−3, 10−4, 10−5, and 10−6 mol/L were employed as the datasets for quantitative analysis. Using the normalized SERS spectral dataset, the deep learning network autonomously searched for features related to quantitative detection under complex conditions with less dependence on Raman peak intensities and without additional preprocessing, which afforded deep-learning-based SERS quantitative detection with excellent reproducibility and feasibility. SERS spectra of stable physical condition were extracted for statistical analysis, and the trained neural network model adequately predicted the trend of variations in the concentration. Using R6G as the probe molecule, a superior recognition result with an accuracy of 98.1 % for the concentrations of 10−3, 10−4, 10−5, and 10−6 mol/L was obtained using a convolutional neural network on the test set. Therefore, this method provides a feasible new strategy to overcome the quantitative detection limitations of current SERS analysis methods.