Application of Computational Methods to Data Integration and Geoscientific Problems in Mineral Exploration and Mining

Thumbnail Image
Trott, McLean R
machine learning , mineral exploration , mining , porphyry copper deposit , epithermal gold deposit , tectonics , signal analysis , continuous wavelet transforms , artificial intelligence
The application of computational methods to data integration and interpretation problems in the mineral exploration and mining space was examined at two distinct scales: deposit-scale studies at the Pimentón and Josemaría epithermal and porphyry copper deposits, and a tectonic examination of the Caribbean at the plate scale using public domain data. Much commonality was found, suggesting that some components of objective workflows are relevant at nearly any scale, with a few modifications. Quality control assessment of data is critical and extends beyond geochemical datasets to any data type required for downstream machine learning processes. Dimensionality reduction, although it may vary in form, is a useful and informative preprocessing step at any scale. Signal analysis techniques like continuous wavelet transform tessellation are relevant to noise reduction and boundary detection for linear data at any scale, when framed appropriately. Examination of workflows for incorporating multiple input feature families was key for this research; again, similarities were found in all cases. Textural metrics were found informative for geological classifications, particularly when involved with geochemical inputs, forming a complementary combination of features with more holistic representation of the rocks in numeric terms, a theme first illustrated at Pimenton and reinforced in Josemaria. Incorporation of distinct datatypes is aided by principal components analysis (PCA) and consideration to spatial coincidence during acquisition and processing of acquired data. Articulation of the concept that classic model metrics like accuracy, precision, and recall are highly influenced by the scale of labelling compared with the scale of feature inputs for many geoscientific problems led to an in-depth examination of this problem during the Josemaría study. The proposed solution, involving the removal of ‘noise’ and generalization of class membership probabilities for linear data predictions, is effective, and incorporates an element of spatial coherency into predictions. The same technique for noise reduction and generalization, continuous wavelet transform tessellation, proved useful on an entirely different scale during examination of hypocenters in the Caribbean region. The method provided an empirical means of locating significant hypocenter density minima along the trend of the arc. Ultimately, it is hoped that this research can facilitate the adoption of semi-automated workflows for large/high-dimensioned datasets or routine tasks and permit human ingenuity to be used on the tasks where creativity is essential.
External DOI