A Novel LIBS-Machine Learning Strategy for Multimetal Detection in Microsized PMMA Particles: Efficient Quantification for Composite Pollution

Microplastics (MPs) have emerged as a critical environmental pollutant, causing composite pollution through their widespread production, usage, and disposal, as well as their capacity to carry other contaminants such as heavy metals. This study presents a novel approach combining laser-induced breakdown spectroscopy (LIBS) with machine learning for the simultaneous quantitative detection of three metals (Cr, Pb, and Cu) in 25 contaminated poly(methyl methacrylate) (PMMA) samples with a diameter of 2 μm. The effects of different preprocessing methods and variable selection techniques on the predictive performance of partial least-squares (PLS) calibration models were investigated. Based on optimized input variables and model parameters, PLS calibration models were developed using mean relative error (MRE), root-mean-square error (RMSE), and coefficient of determination (R2) as evaluation metrics. The models of standard normal variate-competitive adaptive reweighted sampling-PLS (SNV-CARS-PLS) for Cr and Pb, and wavelet transform-CARS-PLS (WT-CARS-PLS) for Cu, demonstrated superior correlation relationships (Cr: Rp2 = 0.9750, Pb: Rp2 = 0.9759, Cu: Rp2 = 0.9088) compared to univariate calibration methods. The values of RMSEp for Cr, Pb, and Cu decreased by 5.495, 9.170, and 3.765 ppm, respectively, while values of MREp decreased by 71.73%, 65%, and 66.81%, respectively. The values of ratio of prediction to deviation (RPD) for three models in -leave-one-out cross-validation (LOOCV) were 20.4, 31.6, and 31.6 respectively. Furthermore, the limits of detection (LODs) for the three heavy metal elements were ≤1.534 ppm. The SNV/WT-CARS-PLS method significantly improved quantitative analysis accuracy, providing essential theoretical and technical support for composite pollution monitoring and prevention in MPs.