A novel approach towards experimental parameters optimization in Laser-induced breakdown spectroscopy

lecture

English

Here we propose a novel and universal method of Laser-Induced breakdown spectroscopy (LIBS) experimental conditions optimization based on machine learning. The simple feedforward neural network (FNN) was trained by empirically measured data. The design of FNN was optimized using a genetic algorithm (GA). As the figure of merit of GA was utilized the signal to noise ratio of selected spectral lines. The input data for FNN can be divided in two groups, one group describing the sample and spectral lines of respective elements (e.g. sample density and hardness, content of selected element, energy levels of selected transitions etc.), and the other group describing the experimental conditions (e.g. laser wavelength and energy, gate delay, gate width etc.). The method is demonstrated and explained in a simple case of single pulse LIBS and two basal parameters – gate delay and laser pulse fluence. Afterwards, we present the optimization for more complex measurement three orthogonal laser pulse (3P LIBS), where the optimization comprises three laser pulse energies, two interpulse delays and gate delay. Finally, we show that the method can work universally even for samples and spectral lines out of the scope of the training data and that with every other measurement the model becomes more precise and robust.

neural network, experimental parameters, LIBS, optimization

PROCHAZKA, D.

21. 9. 2020