Deep learning regression for quantitative LIBS analysis

Abstract

One of the most promising innovation strategies for sorting and recycling post-consumer aluminium scrap is using quantitative Laser-Induced Breakdown Spectroscopy (LIBS) analysis. However, existing methods to estimate alloying element concentrations based on LIBS spectra, such as linear univariate regression and Machine Learning models, are still too limited in their performance to achieve the accuracy demanded by the industry. Therefore, this study presents novel Deep Learning approaches and compares their performance to those of traditional univariate regression and Machine Learning methods in terms of RMSE, MAE, and R2 value. For this evaluation, two sample sets of aluminium pieces are used: one containing 27 certified aluminium reference samples and the second containing 733 post-consumer scrap pieces for which the ground truth concentrations are determined by X-Ray Fluorescence (XRF). Adopting multiple loss functions, one for each element, has proven its significant value for the regression performance. It improves the results for all performance metrics in the Scrap Sample set, and the same is true for the Reference Sample set, except for the coefficient of determination of Fe, Mn and Mg. In addition, the proposed methodology considers the learning prioritisation problem to prevent that learning the concentration of the base element is prioritised over the alloying elements. Although the effect of excluding the base alloy aluminium from the learning is small and not always positive for the performance, demonstrating this effect is also considered valuable. Since the average RMSE on the prediction is just 0.02 wt% for Al and Si, and not more than 0.01 wt% for Fe, Cu, Mn, Mg, and Zn, the best-performing Deep Learning model shows promise for the future of LIBS in metal sorting applications.