Principal component-generalized spectrum-machine learning approach for quantifying gallium in a surrogate plutonium matrix using LIBS

Recent developments in atomic spectroscopy techniques enable rapid quantitative analysis of nuclear material through the implementation of data science techniques. Atomic emission spectra of such materials are often convoluted, owing to their complex makeup and electronic structures. Consequently, performing a chemical analysis using such spectra requires the implementation of advanced analytical methods to understand the relationship between spectral emission features and the chemistry of the material. We present an implementation of spectral analysis methods that combine principal component analysis with supervised machine learning regressions, enabling the quantification of gallium in cerium matrices with superior precision and sensitivity. The proposed principal component-generalized spectrum-machine learning (PC-GS-ML) approach yields marked improvements in model accuracy compared to regressions trained using solely spectral features or PC reduced spectral features. Prediction errors as low as 0.08 wt% Ga are achieved by training boosted ensemble and Gaussian kernel regression models on PC-GS preprocessed laser-induced breakdown spectroscopy data, showing an order of magnitude improvement in the Ga quantification error over prior studies in literature.