Towards real-time calibration-free LIBS supported by machine learning

Calibration-Free Laser-Induced Breakdown Spectroscopy (CF-LIBS) enables multi-elemental quantification without needing standards. This type of approach can be used to analyze complex samples containing traces or gradients of species. This type of diagnosis requires a high level of expertise, and is cumbersome to set up. These constraints limit its application to field diagnostics. Using the MERLIN generalized radiative transfer code, we are able to generate a diversified emission database with no dimensioning limitations. We show that training a convolutional residual network with such a database enables the quantification of 9 species, as well as evaluation of electron density and temperature, without any prior expertise at a rate greater than 10 Hz. The accuracy of this innovative method depends solely on the basic spectroscopic data (emission probabilities and Stark parameters), regardless of the thermodynamic conditions of the laser-induced plasma, as long as it is in Local Thermodynamic Equilibrium (LTE). [Display omitted] •Convolutional Neural Network (CNN) architecture optimized for emission spectra analysis•Physical simulated spectra dataset based on previous WEST tokamak LIBS analysis used for learning and validation procedures•Proof of capability to extract plasma temperature (LTE) and electron density for any spectrum with an accuracy less than 5 %, proof of spectral sensitivity depending the probed species, proof of real-time (10 Hz) CF-LIBS quantification regardless the temperature and electron density within a selected spectral range containing the radiative signature of most of species considered for training.