Implementation of stacked ensemble machine learning for the detection of surrogate plutonium contamination in soil via LIBS

Supervised machine learning methods have demonstrated increased utility for the quantification of lanthanide and actinide elements in atomic spectroscopy applications. This study implements laser-induced breakdown spectroscopy (LIBS) for the identification of plutonium surrogate material (<i>C</i><i>e</i><i>O</i>₂) in soil matrices by training supervised machine learning methods on the recorded spectral data. A bagged ensemble using Random Forest yields the highest sensitivity predictions with a detection limit of 0.015 wt.% <i>C</i><i>e</i><i>O</i>₂. However, high precision in Ce content prediction required the use of a stacked ensemble regression, which provided the superlative Ce quantification model with an error of 0.107% and a detection limit of 0.022 wt.%. The high performance of the stacked ensemble demonstrates its potential to enhance the accuracy and sensitivity of nuclear contaminant detection using field-deployable spectroscopic analyzers in real-world scenarios.