Comparison of Machine Learning Techniques for Mineral Resource Categorization in a Copper Deposit in Peru

The primary objective of this study was to evaluate the effectiveness of three machine learning techniques in the confidence categorization of mineral resources within a copper deposit in Peru: extreme gradient boosting (XGBoost), random forest (RF), and deep neural network (DNN). To achieve this, geostatistical and geometric datasets were employed to categorize mineral resources into measured, indicated, and inferred categories. The dataset included ordinary kriging estimates, kriging variance, average distances, the number of composites, the kriging Lagrangian, and geological confidence. This dataset was used to train the models, followed by the application of smoothing techniques to the initial classification results to ensure a spatially coherent representation of the deposit. The results indicate that the RF model achieved the highest overall accuracy (94%), categorizing 1403.70 million tons (Mt) as measured resources (average grade of 0.43%), 2230.58 Mt as indicated resources (average grade of 0.33%), and 2225.08 Mt as inferred resources (average grade of 0.31%). XGBoost classified a slightly higher tonnage of measured resources (1412.35 Mt) with average accuracy of 91%, while DNN excelled in inferred resources, classifying 2254.64 Mt with accuracy of 93%. Smoothing improved the transitions between categories, reducing discontinuities and providing a more coherent representation of the deposit. The study concluded that machine learning techniques are robust and accurate tools for mineral resource categorization, particularly in geologically complex deposits.