Raman spectroscopy combined with machine learning for the quantification of explosives in mixtures

Detection of explosives and their residues in real time is of paramount importance to homeland security and military. In real-time applications, the suspected materials may contain several chemical compounds making the detection even more challenging. We demonstrate a compact portable Raman spectroscopic tool for quantitative detection of constituent explosives in a binary mixture using machine learning. For the experiment, two samples were considered and mixed at different weight percentages:—1,3,5-trinitroperhydro-1,3,5-triazine (RDX) and ammonium nitrate (AN). Linear regression was employed to quantify the amount of RDX and AN. Regression analyses were conducted using both univariate and multivariate machine learning methods. The Raman spectra were analyzed with and without background correction. Further, various feature/variable selection strategies were explored to find out the best analysis protocol. Our analysis shows that the background correction of the spectra does not improve the accuracy. Among various feature selection techniques, multivariate analysis by considering the total spectra and features associated with only peaks as input gives better results than univariate analysis and multivariate analyses of other sub-spectra. The results demonstrate that Raman spectroscopy combined with machine learning can be used as a reliable, compact, and fast tool for the real-time investigation of explosive mixtures.