Residual compensation algorithm for machine learning-based LIBS quantification optimization

We propose a laser-induced breakdown spectroscopy (LIBS) quantitative prediction algorithm based on residual compensation that can effectively reduce the mean absolute error of prediction (MAEP) and mean relative error of prediction (MREP) by incorporating environmental and sample parameters into the prediction model. A 10-fold cross-validation of the residual compensation model based on four common models, support vector machine regression (SVR), partial least squares regression (PLSR), random forest regression (RFR), and K-nearest neighbor regression (KNNR), was applied to test the aluminum alloy samples. After testing on 10 elements, MAEP and MREP were reduced by an average of 51.8% and 64.8%, respectively, compared with the original PLSR model. For the SVR-based model, the algorithm can reduce MAEP and MREP by 43.0% and 51.1%, respectively.