Quantitative LIBS analysis of microalloyed steel using LASSO–LightGBM modeling with SHAP-based interpretability

Abstract

Microalloyed steels are widely used in the aerospace and automotive industries due to their high strength and corrosion resistance. However, accurately quantifying alloying elements in such complex matrices remains a significant challenge. In this study, laser-induced breakdown spectroscopy (LIBS) combined with machine learning was employed to determine the concentrations of Mn, Cu, Ni, Mo, and Ti in microalloyed steels. High-dimensional spectral data were reduced using the least absolute shrinkage and selection operator (LASSO), followed by regression modeling with univariate calibration, partial least squares regression (PLSR), Support Vector Regression (SVR) and Light Gradient Boosting Machine (LightGBM). Among these models, the LASSO–LightGBM approach achieved the highest predictive accuracy, with R_cv² values consistently reaching 0.996 and the lowest Root Mean Square Error of Prediction (RMSE_P) of 0.0053 wt% for Cu—a 54% reduction compared to PLSR. Furthermore, SHAP analysis directly identified a small number of dominant wavelengths, including covariate features, as principal predictors. Abnormal SHAP responses in specific regions suggested the presence of spectral overlap, highlighting the influence of non-target elements on model outputs. These results clearly demonstrate that integrating LASSO-based feature selection, nonlinear LightGBM modeling, and SHAP-based interpretability yields a robust framework for LIBS-based quantitative analysis, with promising applicability in industrial quality control and materials development.