In Steels Using Laser-Induced Breakdown Spectroscopy

Abstract

The carbon levels in low-alloy steel samples were measured using laser-induced breakdown spectroscopy (LIBS) and a random forest (RF) method. When employing the RF method, the root-mean-square error of cross-validation (RMSECV) criterion was first used to select the spectral range of the spectral variables for RF model input, to prevent over-fitting of the RF model when only a few relevant variables are accompanied by many other variables. Second, the out-of-bag (OOB) error criterion was used to optimize the numbers of decision trees (n_tree) and characteristic variables (m_try) in the RF model, which optimizes the RF structure. The availability of a large amount of relevant spectral information, coupled with the remarkable regression capacity of RF, greatly improved the carbon analytical accuracy. The results showed that the root-mean-square error of prediction (RMSEP) was 0.034 wt.% for the calibration curve method and 0.023 wt.% for the RF method; the reduction afforded by the latter method was 32.4%. Thus, the RF method improved the carbon analytical accuracy for low-alloy steels.