Super-ensemble learning of petrographic features as exploration criteria: Classification of chromitite with unconventional catalysts

Abstract

Sustainable development goals (SDGs) include the conservation and availability of Earth’s resources, and access to clean and affordable energy. Geosciences form the research foundation of Earth resource management, thus necessitating precision and advanced computational modelling for unconventional resource exploration. We developed a super-ensemble learning model that employs the Multinomial Naive Bayes and Extreme Gradient Boosting algorithms to identify exploration targets of naturally occurring catalysts for carbon dioxide methanation. Our model achieved 74% and 70% accuracy for the train and test sets, respectively, with up to 100% higher accuracy than alternative classifiers tested on this dataset. Our work presents a novel application of machine learning for exploring unconventional targets, uncovering the potential of mineral catalysts that require fewer processing steps. It highlights how tree-based, greedy algorithms, easier to implement and interpret than traditional compositional data analysis, can capture complex relationships in a multivariate dataset with high accuracy and translate them into observable characteristics, even directly on a hand specimen. Our work contributes to the advancement of sustainable energy materials which promote resource sustainability, especially in the face of potential noble and critical metal supply shortages by 2050.