Machine learning for quantitative LIBS analysis of aluminum alloys: a comparison of random forest, gradient boosting, and extremely randomized trees

Laser-induced breakdown spectroscopy (LIBS) is a versatile technique for characterizing materials and analyzing elements, but it is limited in its quantitative accuracy due to nonlinear effects. This study uses machine learning (ML) regression algorithms to improve the quantification of commercially pure aluminum and its alloys with multiple alloying elements. A dataset of LIBS spectra from standard samples was used to train three tree-based ML models: Random Forest (RF), Gradient Boosting (GB), and Extremely Randomized Trees (ET). These models were evaluated using standard regression metrics. Prediction errors were consistently low for both in-sample (non-independent) validation and independent validation samples (MAE < 0.25 wt%), confirming strong generalization. The ET model produced the lowest errors, particularly for binary and ternary alloys. The GB model generated the most statistically indistinguishable results compared to the reference values, primarily due to higher prediction variance. The RF model showed intermediate performance. Reliable predictions required at least three representative samples per alloy family, which emphasizes the importance of dataset diversity in LIBS-based machine learning workflows. Therefore, integrating LIBS with ML is an effective strategy for predicting elemental compositions in metallic alloys. Future work will extend this approach to other aluminum alloys and support the development of scalable frameworks for quantitative analysis.