Fingerprint terahertz spectroscopy combined with machine learning for multicomponent mixture analysis

Abstract

The distinctive vibrational features in terahertz (THz) spectroscopy characterize a “fingerprint” of the single-component molecular substance. However, due to componential spectral overlapping and baseline drift, the identification and quantification of multicomponent mixtures are quite challenging for THz spectral analysis. A systematic and feasible strategy has been proposed by combining machine learning with THz spectroscopy for both qualitative and quantitative analysis. After the component number was effectively determined by singular value decomposition (SVD), nonnegative matrix factorization (NMF) and self-modeling mixture analysis (SMMA) were applied to extract componential THz spectra. The difficulties of NMF and SMMA encountered in handling ternary mixtures were solved. The results show component spectra extracted by SMMA are highly consistent with the experimental spectra of pure substances after standardization to correct baseline drift, which greatly facilitates rapid identification of compositions in mixtures. Additionally, compared to back-propagation neural network (BPNN), support vector regression (SVR) predict the contents of each individual component with high robustness and the decision coefficient R2 greater than 0.949. Fingerprint terahertz spectroscopy enhanced by machine learning provided an effective strategy for mixture analysis in practical applications.