Multivariate calibration in Laser-Induced Breakdown Spectroscopy quantitative analysis: The dangers of a ‘black box’ approach and how to avoid them

• Published in: Spectrochimica acta. Part B: Atomic spectroscopy Vol. 144; pp. 46 - 54
• Main Authors: Safi, A., Campanella, B., Grifoni, E., Legnaioli, S., Lorenzetti, G., Pagnotta, S., Poggialini, F., Ripoll-Seguer, L., Hidalgo, M., Palleschi, V.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier B.V 01-06-2018; Elsevier BV
• Subjects: Algorithms; Analytical methods; Artificial Neural Networks; Calibration; Calibration curves, Partial Least Squares Analysis; Cast iron; Laser induced breakdown spectroscopy; Lasers; LIBS; Multivariate analysis; Neural networks; Organic chemistry; Quantitative analysis; Redundancy; Reliability analysis; Software; Software packages; Spectroscopic analysis; Spectroscopy; Artificial Neural Networks; LIBS; Calibration curves, Partial Least Squares Analysis; Cast iron
• ISSN: 0584-8547; 1873-3565
• DOI: 10.1016/j.sab.2018.03.007

The introduction of multivariate calibration curve approach in Laser-Induced Breakdown Spectroscopy (LIBS) quantitative analysis has led to a general improvement of the LIBS analytical performances, since a multivariate approach allows to exploit the redundancy of elemental information that are typically present in a LIBS spectrum. Software packages implementing multivariate methods are available in the most diffused commercial and open source analytical programs; in most of the cases, the multivariate algorithms are robust against noise and operate in unsupervised mode. The reverse of the coin of the availability and ease of use of such packages is the (perceived) difficulty in assessing the reliability of the results obtained which often leads to the consideration of the multivariate algorithms as ‘black boxes’ whose inner mechanism is supposed to remain hidden to the user. In this paper, we will discuss the dangers of a ‘black box’ approach in LIBS multivariate analysis, and will discuss how to overcome them using the chemical-physical knowledge that is at the base of any LIBS quantitative analysis.
[Display omitted]
•The build and optimization of calibration curves for minor elements in cast iron is discussed•Univariate and Multivariate methods are used•The dangers of using multivariate methods as ‘black boxes’ are discussed•Methods are described for assessing the reliability of multivariate approaches•A simple univariate approach may give results equivalent to complex multivariate approaches