Interpreting neural networks trained to predict plasma temperature from optical emission spectra

We explore the application of artificial neural networks (ANNs) for predicting plasma temperatures in Laser-Induced Breakdown Spectroscopy (LIBS) analysis. Estimating plasma temperature from emission spectra is often challenging due to spectral interference and matrix effects. Traditional methods like the Boltzmann plot technique have limitations, both in applicability due to various matrix effects and in accuracy owing to the uncertainty of the underlying spectroscopic constants. Consequently, ANNs have already been successfully demonstrated as a viable alternative for plasma temperature prediction. We leverage synthetic data to isolate temperature effects from other factors and study the relationship between the LIBS spectra and temperature learnt by the ANN. We employ various post-hoc model interpretation techniques, including gradient-based methods, to verify that ANNs learn meaningful spectroscopic features for temperature prediction. Our findings demonstrate the potential of ANNs to learn complex relationships in LIBS spectra, offering a promising avenue for improved plasma temperature estimation and enhancing the overall accuracy of LIBS analysis. ANN can learn spectroscopic trends widely used by domain experts for plasma temperature estimation using emission spectra.