Chemometric methods in plasma spectrometry

V.Dillingerová^1, T. Vaculovič^1, Lukas Nejdl^2, Johannes T. van Elteren^3, Vid S. Šelih^3, Martin
Šala^3, Dino Metarapi^3, V. Kanický^1^

                 1) Faculty of Chemistry, Masaryk University, Brno, Czech Republic

2) Department of Chemistry and Biochemistry, Faculty of agronomy, Mendel university, Brno, Czech
Republic

3) Department of Analytical Chemistry, National Institute of Chemistry,

Hajdrihova 19, Ljubljana, Slovenia


                                      nikadilli@mail.muni.cz,

Plasma spectrometry is frequently used mechanism of multiple analytical methods focusing on
elemental analysis and imaging.  It includes, in particular, mass spectrometry with inductively
coupled plasma (ICP-MS), ICP-MS with laser ablation (LA-ICP-MS) and laser-induced breakdown
spectroscopy (LIBS).  The goal of this study is to explore different approaches for data processing
and statistical analysis of these multivariate data. This work is split into 3 parts; data
reduction for bulk analysis and imaging, image enhancement by convolutional neural networks and
classification of samples.

A data reduction software Ilaps was developed for fast processing and visualisation of mass
spectrometric data. After loading the acquired data, the time-resolved analysis is visualised in an
interactive graph and the peaks and background are automatically selected using a threshold value
or timestamps from the laser ablation system. The program offers multiple processes of analysis,
including despiking, background subtraction, calibration, quantification, data correction and
elemental imaging. To use the software for elemental imaging of LIBS data, a short script was
created to allow import.

Second part of this study focuses on elemental imaging enhancement using fast.ai library [1]. With
destructive methods, such as LA-ICP-MS or LIBS, the sample is ablated by laser beam, resulting in a
pixelated image, where the resolution is dependent among other things, on the beam size. Decreasing
the beam size infinitely is not possible, which leads to the image enhancement by computers. A
ResNet34 model was trained on a dataset of brain images and cellular images. The training set
consisted of original high-resolution images and their low-quality version created by virtual
ablation [2] developed at the National institute of chemistry in Ljubljana. The results were
compared by normalised mean squared error (NRMSE), structural similarity index (SSIM) and peak
signal to noise ratio (PSNR), and for both models, the predicted image ranked better on all 3
metrics in comparison to the ablated image and the image enhanced by bicubic interpolation. The
whole project is open-source and published on GitHub [3].

The last part concentrate on the classification of wine samples using advanced statistical methods
and finding an algorithm for this classification with the best performance. Wines from four
different European countries were analysed using fluorescence spectrometry. The dataset contains
441 samples from Austria, Czech, Serbia and Slovakia. Using their emission spectra, multiple models
were trained and validated: random forest algorithm, k nearest neighbours, linear neural network
and convolutional neural network (CNN). Best performance (97%) was achieved using CNN after
converting spectra to images using gramian angular field.


Acknowledgements:

This work was supported by the Project Grant at MU - MUNI/A/1421/2019 and the ERASMUS+ programme of
the European Union.


[1] Howard, J. and Gugger, S., 2020. Fastai: A layered API for deep learning. Information, 11(2),
p.108.

[2] van Elteren, J.T., Šelih, V.S. and Šala, M., 2019. Insights into the selection of 2D LA-ICP-MS
(multi) elemental mapping conditions. Journal of Analytical Atomic Spectrometry, 34(9),
pp.1919-1931.

[3] V.Dillingerová, Enhancement of the elemental imaging, (2020)
https://github.com/nikadilli/elemental-imaging-enhancement DOI:10.5281/zenodo.3947975