Abstract
Machine learning is a contemporary field that has the potential to supplement the core logging process, expediting and enhancing mineral exploration. Several studies have shown that machine learning may be used with various types of data obtained from the core to identify rock type and alteration, among other features. The purpose of this project was to develop a program that can use XRF data that has been continuously obtained from automated core scanning machines to identify rock type and alteration in core segments. This can be used to help identify key segments of core that may need more attention from a geologist to confirm indications of potentially economical mineralization.
This program was developed in Python and evaluated using data from 32 boreholes in South Australia, which are publicly available on the Minalogger platform. The random forests, gradient boosting trees, and support vector machines algorithms were evaluated with a sample set of 6 boreholes, and random forests was found to be the superior algorithm in terms of overall accuracy, precision, and recall. As such, the random forest algorithm was used in the program. The rock types that were used for identification included conglomerates, sandstones, volcanic rocks, fine-grained sediments, skarns, plutonic rocks, metamorphic rocks, and veins.
The final program was produced and trained with 75% of the core data used for training and 25% used for testing. This resulted in a 97% accuracy rate, with vein identification having a precision of 89% and a recall of 59%. The veins were most commonly mistaken for igneous and metamorphic rocks, which may be mitigated in the future by using imagery along with the XRF data. These results demonstrate that the random forest algorithm can be used to help in the identification of rock type and alteration in rock core.