Component identification for Raman spectra with deep learning network

Abstract

Raman spectroscopy is widely used in the research of the molecular structure of substances because of the advantages of no invasion, no damage and no interference from water. Meanwhile, component identification for mixtures is still challenging in Raman spectra. In this paper, a graphics-based sample-generating method and a model based on deep-learning for component identification was proposed. Convolution neural network (CNN) model is an important part of deep learning network and CNN models was utilized to assess the possibility of the presence of components in samples. As is shown in the comparative studies, the model was sensitive to the relative position of the characteristic peaks and could learn spectra features in mixtures. The deep-learning based component identification method showed more robustness than conventional linear fitting methods. Therefore, the method provided a valid approach to component identification for mixtures and has the potential in spectra component analysis.