Machine learning in laser-induced breakdown spectroscopy as a novel approach towards experimental parameter optimization

• Published in: Journal of analytical atomic spectrometry Vol. 37; no. 3; pp. 63 - 612
• Main Authors: Prochazka, David, Po ízka, Pavel, Hruška, Jakub, Novotný, Karel, Hrdli ka, Aleš, Kaiser, Jozef
• Format: Journal Article
• Language: English
• Published: London Royal Society of Chemistry 09.03.2022
• Subjects: Alloying; Aluminum base alloys; Artificial neural networks; Data points; Laser induced breakdown spectroscopy; Lasers; Learning theory; Line spectra; Low alloy steels; Machine learning; Mechanical properties; Noise prediction; Optimization; Parameters; Physical properties; Signal to noise ratio; Spectrum analysis; Training
• ISSN: 0267-9477; 1364-5544
• DOI: 10.1039/d1ja00389e

Similar to other analytical techniques, the performance of laser-induced breakdown spectroscopy (LIBS) is significantly influenced by the selection of optimal experimental parameters. The optimization of LIBS is challenging because the laser-matter interaction and subsequent plasma formation are influenced not only by selected experimental parameters but also by the physical and mechanical properties of the sample. The goal of this work is to develop an artificial neural network (ANN) that is able to predict the signal-to-noise ratio (SNR) of selected spectral lines based on specific experimental parameters (laser pulse energy and gate delay) and on the sample's physical and mechanical properties. The ANN training was based on input data obtained from a high number of measurements of three certified materials with highly different mechanical and physical properties (low alloyed steel, glass, and aluminium alloy) with 2079 combinations of experimental parameters - gate delay (GD) and laser pulse energy ( E ). The ANN was optimized in terms of the number of neurons and hidden layers. The minimal number of input data points was studied with emphasis on the ANN prediction accuracy expressed as the determination coefficient R 2 (predicted vs. measured values). The number of input data points was studied from three points of view - a minimal number of experimental parameters for one matrix, a minimal amount of data from different matrices, and a minimal number of different spectral lines. It has been shown that at least 20 different combinations of experimental parameters are necessary for one matrix to obtain reasonable performance of the ANN. However, only ten combinations are needed when a new matrix is added to the working model. It has also been shown that the prediction accuracy is poor for spectral lines which were not part of the training data. Finally, the ANN was utilized to predict the SNR of selected spectral lines in a specific range of experimental parameters. The parameters with the maximal SNR were studied, and the values were discussed with an emphasis on sample properties. It has been concluded that the optimization process can be substituted or significantly shortened by means of the ANN. Samples with different mechanical and physical properties were measured by LIBS under diverse experimental conditions. The results were used to train a neural network. By means of the neural network, the optimisation process was significantly reduced.