Uso da técnica laser-induced breakdown spectroscopy (LIBS) para análise direta de ligas metálicas: estratégias de normalização, calibração univariada e multivariada e modelos de classificação

Abstract

This academic master's dissertation was devoted to the development of analytical methods for the determination of Al, Cr, Cu, Fe, Mn, Mo, Ni, Ti, V and Zn in alloys and steels. The main purpose of the study was to present the Laser-induced Breakdown Spectroscopy (LIBS) as a viable alternative for the direct analysis of alloys and steels using chemometric tools to interpret the obtained data. Initially, the optimization of the parameters of the LIBS equipment was done using Doehlert design, varying the laser energy in 7 levels (30 to 80 mJ), delay time in 5 levels (0 to 2 μs) and spot size in 3 levels (50 to 150 μm). The chosen compromise condition was 60 mJ of energy, 0.9 μs of delay time and 100 μm of spot size, which were applied to 80 samples. The reference values of the analytes were obtained using the X-ray Fluorescence (XRF) technique for the construction of calibration models.To minimize signal variations and sample matrix differences, twelve normalization modes were tested and two calibration strategies were studied: multivariate calibration using Partial Least Squares (PLS) and univariate calibration using area and height of several emission lines. Thus, we search to identify the best mode of normalization, emission line and calibration strategy for each analyte. For most analytes, there was no significant difference between the normalization modes and also between the univariate and multivariate calibration. Classification models were applied to identify the samples in 3 different groups. K-nearest neighbor (KNN), Soft independent modeling of class analogy (SIMCA) and Partial-least squares-discriminant analysis PLS-DA were used in 3 different matrices: concentrations obtained using XRF, height and area of the LIBS emission lines (total of 57 emission lines). When comparing the models, some merit figures were evaluated, such as accuracy, sensitivity, false alarm rate and specificity. The classification model that obtained the best results was KNN. As a conclusion of the work, factorial design was useful to obtain an adequate analysis condition for all analytes and samples simultaneously, saving time and resources. Normalization modes were effective to minimize signal variations and differences in sample matrices. Univariate models were more satisfactory than multivariate models. In the case of classification models, it was possible to identify the samples, being the KNN model more efficient than the others.