Correlation-based carbon determination in steel without explicitly involving carbon-related emission lines in a LIBS spectrum.

ABSTRACT

As any spectrochemical analysis method, laser-induced breakdown spectroscopy (LIBS) usually relates characteristic spectral lines of the elements or molecules to be analyzed to their concentrations in a material. It is however not always possible for a given application scenario, to rely on such lines because of various practical limitations as well as physical perturbations in the spectrum excitation and recording process. This is actually the case for determination of carbon in steel with LIBS operated in the ambient gas, where the intense C I 193.090 nm VUV line is absorbed, while the C I 247.856 nm near UV one heavily interferes with iron lines. This work uses machine learning, especially a combination of least absolute shrinkage and selection operator (LASSO) for spectral feature selection and back-propagation neural networks (BPNN) for regression, to correlate a LIBS spectrum to the carbon concentration for its precise determination without explicitly including carbon-related emission lines in the selected spectral features.