Calibração multivias em Laser-Induced Breakdown Spectroscopy (LIBS) para determinação de Ca, K e Mg em amostras de suplemento mineral para humanos

Abstract

This work intends to present a new approach for modeling LIBS data resolved in time using multiway algorithms, aiming to achieve the advantage of second order. The algorithms evaluated in this study were: Multivariate Curve Resolution- Alternating Least Squares (MCR-ALS), Parallel Factor Analysis (PARAFAC) and Unfolded Partial Least-Squares with Residual Bilinearization (U-PLS/RBL). The analytes monitored in this work were Ca, K and Mg in samples of mineral supplement. This study was divided into four (4) parts. In Part 1 the validation of the algorithms was performed using samples of mixtures of the salts and bases of the studied analytes. In parts 2 and 3, mineral supplement samples were used to assess the predictive capacity of the models, with 6 samples in the calibration set (Part 2) and 17 samples in the calibration set (Part 3). Finally, conventional univariate models were built and the results obtained compared with the performance of the multivariate algorithms (Part 4). The results obtained in the first part demonstrated a good performance for the evaluated algorithms, with emphasis on the U-PLS/RBL. The values obtained for RMSEV by MCR-ALS were 3.4% pp-1, 0.4% pp-1 and 0.9% pp-1 for Ca, K and Mg, respectively. For PARAFAC the values obtained for Ca, K and Mg were 11% pp-1, 0.3% pp-1 and 1.1% pp-1, respectively. The values obtained by U-PLS/RBL were 2.5% pp-1 for Ca, 0.2% pp-1 for K and 1.2% pp-1 for Mg. As for the results obtained in Part 2, a decrease in the performance of the MCR ALS and PARAFAC algorithms was observed when the supplement samples were used as test samples, with high values of LoD and LoQ by both models for the three monitored analytes. The U-PLS/RBL showed an excellent performance, with low values of LoD and LoQ for Ca (0.4% pp-1 and 1.2% pp-1, respectively) and K (LoD: 0.1% pp-1 and LoQ: 0.2% pp-1), while for Mg some samples were below or close to loD (1.0 pp-1) and LoQ (3.0 pp-1). With the inclusion of more samples in the calibration set (Part 3) the performance of the MCR-ALS and PARAFAC algorithms improved for some analytes, however the LoD and LoQ values still remained high for the three monitored analytes. Regarding the U-PLS / RBL, there were no significant improvements for Ca and K, while for Mg a certain improvement in the values of LoD and LoQ was observed, 0.2 pp-1 and 0.7 pp-1, respectively. Conventional univariate calibration (Part 4) did not achieve good results, which was expected, since selective lines are necessary to obtain good results. Therefore, these results confirm that the U-PLS / RBL model was able to predict the analytes of interest in unknown samples, being possible to achieve the second order advantage.