Analysis of copper content combined laser-induced breakdown spectroscopy with feature selection

In laser-induced breakdown spectroscopy, principal component analysis (PCA) and partial least squares regression (PLSR) can achieve feature dimensionality reduction by removing redundant information while retaining the overall useful information of the original features. However, these linear feature selection methods fail to convey the specific meanings of the selected variables from the original spectrum. To understand the physical significance of analyzing spectral lines effectively and enhance the accuracy of quantitative analysis, Elastic Net and Lasso were utilized to remove redundant spectral lines for determining the copper content in LIBS. Spectral analysis was conducted on 21 kinds of semi-artificial copper ore samples, identifying 11 atomic spectral lines and 18 ionic spectral lines for modeling analysis. When α = 0.904 and λ = 0.269, the Elastic Net was created with the best fitting performance. The values of coefficient of determination R² and RMSE are 0.9798 and 1.049 respectively. After removing redundant spectral information, the regression coefficient matrix reveals only 7 atomic spectral lines and 2 ion spectral lines remaining in Elastic Net. For the prediction capability, the mean square error (MSE) and mean absolute percentage error (MAPE) are 1.1971 and 8.21%, respectively. Compared with PLSR, the RMSE and MAPE of the proposed Elastic Net decreased by 24.92% and 35.15%, respectively. This work provided a novel chemometrics method to reduce the redundant lines on quantitative analysis in LIBS applications.