Near-the-Line Steel Slag Analysis Using Laser-Induced Breakdown Spectroscopy: Traditional Univariate Versus Machine Learning Calibration Methods

This work is focused on rapid quantitative analysis of slag in the steel industry for improved process control. The novel approach in this work is a direct comparison of two methods to calibrate and quantify spectral data from the slags. Calibration was first done with the most prevalent method in quantitative optical emission spectroscopy (OES) of solids, the univariate ratio method. The second method is an advanced multivariate analysis (MVA) algorithm termed Elastic Net, allowing to include several lines for each element in the calibration functions. In both methods, the output is mass fraction ratios of the analyte element (or compound) to a matrix element (compound). The actual mass fractions of each compound are calculated by sum normalization assuming the matrix to make up the difference up to 100%. The metric used to evaluate the performance of the methods in terms of accuracy is the parameter σ rel calculated as the ratio of the root mean square (RMS) deviation from values obtained by X-ray fluorescence (XRF) divided by the average mass fraction of the compound, expressed in percent. A bit surprising, the main outcome of the comparison is that there is very little difference in the performance of the two methods. One exception is the analysis of MgO, where the elastic net gives significantly better accuracy. Presumably, this is due to the use of multiple lines for Mg to build the calibration function. This is very encouraging, since MgO is a major compound in most slags that needs to be determined accurately. It is suggested to improve accuracy further by means of separate calibrations for a limited number of slag types.