Transfer learning improves the prediction performance of a LIBS model for metals with an irregular surface by effectively correcting the physical matrix effect. Issue 7 (24th May 2021)

• Record Type: Journal Article
• Authors: Shabbir, Sahar; Zhang, Yuqing; Sun, Chen; Yue, Zengqi; Xu, Weijie; Zou, Long; Chen, Fengye; Yu, Jin
• Is Part Of: Journal of analytical atomic spectrometry. Volume 36:Issue 7(2021)
• Page Start: 1441
• Page End: 1454
• Publication Date: 2021-05-24
• Subjects: Atomic spectra -- Periodicals; Atomic absorption spectroscopy -- Periodicals; 543.0858
• DOI: 10.1039/d1ja00076d
• Languages: English
• ISSNs: 0267-9477

Abstract

Abstract : Transfer learning allows the effective corrections of both chemical and physical matrix effects in such a way that metals with irregular surfaces can be directly analyzed by LIBS for an accurate elemental determination. Abstract : This work was designed to observe and further correct the physical matrix effect in the analysis of solid materials with laser-induced breakdown spectroscopy (LIBS), an effect that arises when a calibration model established with a set of certified reference materials (CRMs) with a smooth surface is used for prediction with LIBS spectra acquired from materials with an irregular surface, such as scraps produced by an automobile shredder for example. CRMs of an aluminum alloy were prepared in such way that one half of the surface of each sample was mechanically destroyed to present an asperity similar to a scrap. LIBS measurements were then performed in the two halves of the sample surface in the same conditions. The spectra acquired from the smooth part of the sample surface served as a training data set to train calibration models, while those taken from the irregular part of the surface were used as a validation data set to assess the prediction performance of the calibration models. Four elements (magnesium, silicon, iron, and zinc), which are important for the recycling of aluminum alloy scraps, were analyzed. The study started with univariate models, where the influence of the surface asperity was clearly observed. A first correction with internal standard normalization showed limited effectiveness. Further improvements were attempted using a machine learning-based multivariate regression, where the calibration performance of the models was significantly improved thanks to an optimized correction of the chemical matrix effect, whereas the prediction performance was still unsatisfactory due to the surface asperity alternation of the validation samples. A transfer learning-based regression model, where a part of the irregular samples joined the training data set, was thus developed to effectively correct both chemical and physical matrix effects and allowed significantly improved the performances of both calibrations, with a relative error of calibration (REC) at 2.3%, and prediction for irregular samples, with a relative error of prediction (REP) at 16.3% on average for the 4 tested elements.